The mRNA decay rates of the transcripts of 74 highly disordered proteins and 536 highly ordered proteins were compared. The mRNA decay rates for the transcripts encoding highly disordered proteins (0.190871 h^-1) are more than twice that observed for the transcripts encoding highly ordered proteins (0.084944 h^-1) (Figure 1b). A statistically significant difference (P < 0.02) between mRNA decay rates for transcripts encoding highly ordered and highly disordered proteins was found, with the highly disordered datasets having higher mRNA decay rates. The mRNA decay rates for the transcripts encoding 1,980 disordered proteins (0.177596 h^-1) and 1,858 ordered proteins (0.096878 h^-1) were also compared and a similar trend was obtained (Figure S1b in Additional data file 1).

We divided the 33,869 proteins into bins by percentage of disordered residues. When we compared the mRNA decay rates for each of the bins (Figure 2b), there was no significant difference between them. Although this result does not suggest that all disordered proteins show a significant association with higher mRNA decay rates, it does concur with our previous analysis of the (highly) ordered and (highly) disordered protein datasets, in showing a distinct difference between mRNA decay rates for both groups.

The protein stability measures of the highly disordered (179) and highly ordered groups (1,396) were also compared. We found a significant difference (P < 0.0005) between the half-lives of highly ordered and highly disordered proteins, with highly disordered proteins having longer half-lives (Figure 1c).

Consistent with our analysis of decay rates, we divided the 8,666 disordered proteins into bins by percentage of disordered residues. Protein stability indices showed no significant affiliation to a particular binned group, although the (80,100]% disorder bin showed much higher half-lives than the other binned groups (Figure 2c).

Since trends were observed between both mRNA decay rate and disorder, and protein half-life and disorder, the half-lives and decay rates were also compared to see if a relationship existed between mRNA decay rate and protein half-life. The Pearson correlation value between 1,446 overlapping sequences (-0.06) was not significant and suggested that these two characteristics are independent.

Approximately one-quarter of protein coding transcripts are predicted miRNA targets (Table 2). The proportion of transcripts encoding highly disordered proteins that are likely to be miRNA targets is approximately twice that of transcripts encoding highly ordered proteins (Figure 1d; Table 2). The frequency of transcripts with at least one predicted miRNA target site is over-represented in the transcripts encoding highly disordered proteins (P < 0.003) and under-represented in the transcripts encoding highly ordered proteins (P < 0.00001) compared to all transcripts together (Figure S2a in Additional data file 1). A similar trend is observed when comparing the datasets of transcripts encoding ordered and disordered proteins (Table 2); the proportion of the transcripts encoding disordered proteins that are predicted as miRNA targets is approximately twice that of the transcripts encoding ordered proteins (Figure S1c in Additional data file 1; Table 2). miRNA targets are over-represented in the transcripts encoding disordered proteins (P < 0.00001) and under-represented in the transcripts encoding ordered proteins (P < 0.00001) compared to all transcripts together (Figure S2b in Additional data file 1).

For the transcripts encoding the proteome, the percent likely to be targeted by miRNA ranges between 13.2% and 37.6% (Figure 2d; Table 2). The percentage of transcripts regulated by miRNA increases (approximately 8%) with increasing percentage of protein disorder for the first three binned categories (Figure 2c; Table 2). The percent of predicted miRNA targets for transcripts remains high (35.1%) for the (60,80]% disorder category and low (13.2%) for the 80100% disorder category. Consistently, the likely miRNA targets are under-represented in the [0,20]% and (80,100]% disorder categories at P < 0.00004 (Figure S2c in Additional data file 1) and over-represented in the remaining three classes (P < 5.8 × 10^-7; Figure S2c in Additional data file 1).

Similar trends are obtained using the PicTar (4-Way and 5-Way) software 4346 (Figures 1d and 2d; Figure S1c in Additional data file 1). The trends were not observed using mirBase 51 and this could be because this prediction algorithm is reported to have a higher false positive rate than the other two programs (PicTar and TargetScanS) 474950. Redundancy in the datasets makes very little difference to the outcome (Table S3 in Additional data file 2). For example, the proteome and the protein sets filtered for redundancy have very similar percentages of transcripts predicted as targets of miRNA (Table 2; Table S3 in Additional data file 2).

We investigated the patterns of the predicted miRNA targets in the transcripts for disordered proteins in relation to the different expression profiles (Figures 3 and 4 and Table 3). The probes on the microarray chip have a higher representation of predicted miRNA targets (38%) in comparison with the transcriptome encoding the human proteome (26.45%) (Table 2). We compared the protein coding transcripts for the five datasets (Figure 3) using the probes on the microarray chip as a universal protein baseline. The data from the constitutive group had too few data points from which to make inferences (Table 3 and Figures 3 and 4). The tissue-specific expressors (gold) and the high expressors (dark blue) have high expression levels. The main difference between the two classes is that the tissue-specific expressors (gold) have high expression in one or few tissues (Figure 3) and the high expressors (dark blue) have high expression in almost all tissues (Figure 3). These two groups characterized by high levels of gene expression have high percentages of transcripts predicted as miRNA targets (68.09% and 65.85%, respectively; Table 3 and Figure 4). The medium expressors (green) and the low or transient expressors (white) with more moderate levels of gene expression have lower percentages of predicted miRNA targeting (48.39% and 48.06%, respectively). These results suggest that the transcripts of disordered proteins with high levels of expression are more likely to be regulated by miRNA compared to those with moderate and low or transient expression. In addition, the transcripts of highly disordered proteins belonging to the four expression profiles (tissue-specific, high expressors, medium expressors and low or transient expressors) are more likely to be miRNA targets than the transcripts on the microarray chip (Figure 4b). This observation supports the trend observed previously (Table 2) that transcripts encoding disordered proteins are more likely to be targeted by miRNAs compared to protein coding transcripts in general (Figure 4; Figures S1c and S2c in Additional data file 1).

To our knowledge, this study presents the first estimate of the percentage of proteins of the human proteome with at least one predicted ubiquitination site and the percentage of residues predicted as ubiquitination sites. We predict that 70.71% of proteins have at least one ubiquitination site and 0.42% of amino acid residues in the proteome are ubiquitination sites.

The percentage of proteins predicted to contain at least one ubiquitination site and the percentage of residues predicted as ubiquitination sites are higher in disordered proteins compared to ordered proteins. Comparing the highly disordered proteins with the highly ordered proteins, we observe increases of 33.81% and 42.50% in the percentage of proteins possessing at least one ubiquitination site and the percentage of residues predicted to be ubiquitination sites, respectively (Figure 1e). The proteins possessing at least one ubiquitination site are slightly over-represented in the highly disordered proteins (P < 0.98; Figure S3a in Additional data file 1) and grossly under-represented in the highly ordered proteins (P < 2.2 × 10^-16; Figure S3a in Additional data file 1). The first trend is not statistically significant. The predicted ubiquitination sites are over-represented in the highly disordered proteins (P < 2.2 × 10^-16; Figure S4a in Additional data file 1) and under-represented for the highly ordered proteins (P < 0.002; Figure S4a in Additional data file 1). Comparing the disordered proteins with the ordered proteins, we observe increases of 33.57% and 12.8% in the percentage of proteins possessing at least one ubiquitination site and the percentage of residues predicted to be ubiquitination sites, respectively (Figure S1d in Additional data file 1). Proteins with one or more predicted ubiquitination sites are over-represented in the disordered datasets (P < 2.2 × 10^-16; Figure S3b in Additional data file 1) and under-represented in the ordered proteins (P < 2.2 × 10^-16; Figure S3b in Additional data file 1). A similar trend is obtained for the percentage of residues predicted as ubiquitination sites.

The relationship between the percentage of proteins with at least one ubiquitination site and the percentage of protein disorder is complex and non-linear, while the percentage of residues predicted as ubiquitination sites and the percentage of protein disorder are positively correlated. The percentage of proteins predicted to have a ubiquitination site increases with the percentage of protein disorder for the first three disorder categories (Figure 2e). The percentage of proteins predicted to have a ubiquitination site remains high at 74.3% for the (60,80]% disorder class and then drops significantly to 55.8% for the (80,100]% disorder category. This is consistent with proteins with one or more predicted ubiquitination sites being over-represented in the (20,40]%, (40,60]% and (60,80]% disorder categories (P < 0.04; Figure S3c in Additional data file 1) and under-represented in the [0,20]% and (80,100]% disorder categories (P < 0.00005; Figure S3c in Additional data file 1). On examination of the second ubiquitination descriptor, a different trend is observed; the percentage of residues predicted as ubiquitination sites increases as the percentage of protein disorder increases, illustrating a strong positive correlation between the two variables (Figure 2e). Proteins with one or more predicted ubiquitination sites are under-represented in the [0,20]% disorder category and over-represented in the remaining four disorder classes (P < 2.2 × 10^-16; Figure S4c in Additional data file 1).

As lysine is over-represented in disordered regions 17677, we investigated the percentage of residues predicted as ubiquitination sites in relation to the percentageof protein disorder, taking into account lysine residue biases (Figure S5a in Additional data file 1). First, we calculated a correlation coefficient for the percentage of predicted ubiquitination sites and the percentage of lysine composition for the five disorder categories and obtained a strong positive correlation (R = 0.844772). Second, we normalized the number of predicted ubiquitination sites with respect to the number of lysines for each dataset. The trends observed for the percentage of predicted ubiquitination sites normalized for lysine frequency and disorder are similar to those obtained with the percentage of predicted ubiquitin sites and disorder ignoring lysine biases (Figure S5b in Additional data file 1). Comparing the disorder categories with the order categories, the calculations normalized using the lysine frequency result in differences that are smaller in magnitude. For example, comparing the highly disordered proteins with the highly ordered proteins, an increase of 23.5% is observed instead of 42.5%, and comparing the disordered proteins with the ordered proteins, an increase of 4.4% is observed instead of 12.8%.

On consideration of the gene expression levels for transcripts encoding disorder in proteins, two main trends emerge. Firstly, on average, the transcripts encoding disordered proteins are expressed to a significantly lower extent than those encoding ordered proteins. This suggests that, typically, the cell has evolved regulatory mechanisms to ensure that it does not have a large proportion of highly expressed transcripts encoding high amounts of protein disorder. Secondly, for the highly disordered proteins, there are five broad classes of gene expression patterns observed. These are constitutive expressors, high expressors, medium expressors, tissue-specific expressors and low or transient expressors. The constitutive expressors represent the transcripts that are constitutively and highly expressed across all samples. The high expressors contain protease inhibitors (specifically cysteine proteases) and enzyme regulatory functions. These proteins function in order to prevent their enzyme counterparts from cleaving peptide bonds; it is thus expected that their expression levels should remain relatively high in unstimulated tissue samples. This group is also enriched in functions directly related to splicing machinery, DNA packaging, nucleosome assembly and the components involved in DNA-protein and protein-protein complex assembly. In the medium expressor group, the transcription factors are predominantly positive regulators of transcription and are involved in muscle, cellular differentiation and general tissue development processes. The tissue-specific expressors are enriched in the transcription factors that target the nuclear hormone receptors and the ligands that are coordinately regulated with their receptor binding partners. The tissue-specific expressors are predominantly negative regulators of cell organization and promote complex disassembly and DNA unwinding and replication.

These novel observations have high biological relevance. Disordered proteins have important roles in the cell 12345678910111213141516171819202122232425262728293031; naturally they have to be expressed to carry out their specified function, but high levels of highly disordered proteins in the cell can cause a major problem through protein misfolding, misidentification and mis-signaling 24. Our analysis suggests that one of the ways in which the cell keeps the level of highly disordered proteins under control is to keep the expression levels of transcripts encoding them low.

A recent study by Paliy and colleagues 78 claims that protein disorder is weakly positively correlated with gene expression in Eschericha coli. In our study, however, the trends between protein disorder and gene expression are complex and non-linear. Differences exist between the trends observed in E. coli 78 and those we report for human data. These differences are attributed to the differences in the methodologies (such as the disorder prediction methods and the definition of gene expression levels) and the consideration of species from disparate taxonomical classes. It is widely accepted that E. coli (a prokaryote) has about 5% of proteins that contain disordered regions whilst human (a eukaryote) has approximately 30% 6. Paliy et al. 78 examined the highly expressed transcripts for proteins possessing high levels of predicted disorder in a prokaryote (E. coli). The types of genes that fall into this category encode RNA and protein chaperones, protein carriers, transcriptional and translational regulators and multi-enzyme complexes. Some of the genes are found only in prokaryotes - these include the peptidoglycan-associated lipoprotein and the glycine cleavage Complex H protein 78 - whilst other genes exist in all taxonomical classes and some of these are identified in our study (for example, the ribosomal proteins and the translational initiation factor). The products of these transcripts whilst highly disordered are required by all cell types and by all species (prokaryote or eukaryote) and the transcripts are constitutively expressed at high levels. In contrast to the E. coli study 78, we find distinct transcripts encoding highly disordered proteins that are low or transiently expressed. These proteins are predominantly transcription factors or activators of transcription involved in developmental processes specific to complex higher organisms (Figure 3).

We examined trends between mRNA decay rates and amounts of disordered residues, and between protein half-lives and frequency of disordered residues. The mRNA decay rates for the transcripts encoding highly disordered proteins are more than twice that observed for the transcripts encoding highly ordered proteins. Large quantities of proteins containing high levels of disordered residues can cause function problems for the cell 24 and a higher mRNA decay rate could indicate the necessity for removal of potentially problematic proteins. Our finding of a correlation between high mRNA degradation rates and disordered proteins would appear to be in agreement with this.

The correlation between increasing amounts of protein disorder and longer half-lives was not expected since intrinsically disordered sequences are known to be extremely susceptible to proteolytic degradation 79. However, these results were consistent with findings reported in Yen et al. 75, who observed an enrichment of disorder promoting residues in more stable proteins. This may be a feature of the way in which stability was measured. Attachment of an amino-terminal GFPS tag to a protein in the global protein stability assay may interfere with cellular localization and the authors of this study recognize that the stability values tended only to be reliable for nuclear proteins. Additionally, it is likely that this tag affects correct folding of the protein and might obscure amino-terminal degradation signals (N-degrons), which are a major determinant of stability in eukaryotic sequences 80. Considering these features of the available dataset, it may be that proteins with longer half-lives are enriched in the set that coincided with our data. However, the occurrence of highly stable sequences with long half-lives observed within sequences containing between 80% and 100% of disordered residues correlates well with the hypothesis that highly disordered proteins exist as complexes in vivo. A similar conclusion was drawn in the global protein stability profiling study where Yen et al. suggest that this mechanism constitutes a protection mechanism from cellular protein degradation machinery 75.

Recent work on protein disorder 8182 arrived at similar conclusions to our study. They suggested that certain disordered proteins may be required to remain in the cell for long periods of time, and thus need to avoid the degradation process. They suggest that such avoidances are evident by an increase in protein stability for some disordered proteins. Separately, they also found a correlation between decay rates and mRNA stability for disordered proteins, in agreement with our analysis.

We find a significantly higher level of predicted miRNA regulation of the transcripts encoding highly disordered proteins compared with the transcripts encoding highly ordered proteins. The predicted levels of miRNA regulation of the transcripts encoding highly disordered proteins are twice that observed for the transcripts encoding highly ordered proteins. Over one-third of the transcripts encoding disordered proteins are predicted to be regulated by miRNA. One-fifth of the transcripts encoding ordered proteins are predicted to be miRNA targets. Furthermore, miRNA-regulated gene expression is over-represented in transcripts encoding disordered proteins and under-represented in the transcripts coding for ordered proteins. These trends take into account 99% of the transcriptome encoding disordered proteins (that is, excluding the (80,100]% disorder category). We find that the transcripts of highly disordered proteins with high levels of expression are more likely to be affected by miRNA compared to those with moderate and low or transient expression. We provide strong evidence for miRNA regulation being particularly important for transcripts encoding disordered proteins. The observations make sense in a biological context. Typically, if a protein has high proportions of disorder, it is rapidly degraded in the cell 32. For the cell, it would make economic sense to have an analogous system in place to handle the flagging up and degradation of the corresponding mRNA at the transcriptome level. This increased likelihood of miRNA binding to mRNA molecules that encode disordered proteins would regulate the gene expression of the mRNA molecule and prevent undesirable and wasteful translation of proteins no longer required by the cell.

The percentage of residues predicted as possible ubiquitination sites increases with increasing amounts of disorder. Interestingly, the relationship between these two properties is linear and positive. The trend between the percentage of proteins predicted to have one or more ubiquitination sites and disorder is more complex. The (80,100]% category has the lowest proportion of proteins predicted to have one or more ubiquitination sites compared to the remaining four categories. This follows a similar trend observed for mRNA decay rates and predicted miRNA targeting. This suggests that a significantly lower proportion of these proteins are likely not to be down-regulated by ubiquitination. This supports the observation described earlier that a significant proportion of highly disordered proteins is required to be expressed at high levels in all tissues or some tissues, and some are sometimes constitutively expressed. If high proportions of these highly disordered proteins and their corresponding transcripts did have positive signals for targeted degradation, this could adversely affect fitness. Additionally, highly disordered proteins with one or more predicted ubiquitination sites that are not constitutively or highly expressed may have a higher chance of being removed from the cell, as they are likely to have a higher density of ubiquitination sites. The (80,100]% disorder category is most likely to have the highest density of ubiquitination sites (Figure 2e).

The increase in the decay rate of the transcripts encoding disordered proteins is likely attributable, in part, to the increase in predicted miRNA regulation. The transcripts encoding disordered proteins are targeted to a higher extent by miRNA compared to the transcripts encoding ordered proteins. This will result in the down-regulation of gene expression. The absolute gene expression levels and the predicted miRNA regulation are anti-correlated. The overall decrease in the gene expression of the transcripts encoding disordered proteins is likely attributable, in part, to the increased miRNA targeting that results in the down-regulation of these transcripts.

On the one hand, the majority of the disordered proteins have evolved with higher mRNA decay rates, higher levels of miRNA targeting, and higher levels of ubiquitination, which overall result in lower gene expression levels and protein levels for a high proportion of these disordered proteins compared to the ordered proteins. On the other hand, it is shown that for a significant proportion of highly disordered proteins, the converse is true. For the (80,100]% disorder class there is a decrease in mRNA decay rates, lower proportions of miRNA targeting and lower proportions of proteins being targeting for ubiquitination. These properties play a role in the high levels of gene expression observed in the highly disordered proteins compared to proteins with less disorder. The regulation of disordered proteins is affected by the various factors studied, and the relationships between these properties and protein disorder are inter-related, non-linear and complex.

Chen et al. 83 performed a structural biology analysis for the purpose of studying associations between structural vulnerability and co-expression in yeast and human. They claim that structural vulnerability (structural disorder) affects gene co-expression in a quantifiable manner 83. In their study, they consider post-transcriptional regulation of transcripts of highly vulnerable proteins and find that 45% of human genes are predicted to have a least one miRNA target site compared to 82.9% of extremely vulnerable genes (87 out of 105) 83. The mean number of miRNA target sites is 2.66 for human genes and 6.01 for vulnerable genes 83. They show that vulnerability (disorder) requires significant additional regulation at the post-transcriptional level. This is an observation also made in our study; however, our miRNA study provides a more comprehensive analysis. For example, we investigate three different views of disorder. The first definition divides the transcriptome into transcripts encoding disordered proteins or ordered proteins. Second, transcripts encoding highly disordered or highly ordered proteins are considered. Third, we examined miRNA targeting of mRNAs encoding proteins with different percentages of disordered residues. Our study therefore provides more information in relation to disordered and highly disordered proteins. For example, we find that transcripts encoding disordered proteins are more likely to be targeted by miRNA than the transcripts encoding highly disordered proteins (Figures 1d and 2d; Figure S1c in Additional data file 1). Additionally, the method and the materials of Chen et al. 83 provide an estimate of miRNA targeting higher than our analysis and higher than estimates in other studies 3945. Our overall estimate of 26.5% for miRNA targeting agrees well with other estimates for miRNA targeting of protein coding transcripts 3945. miRNAs are shown to preferentially target genes with high transcriptional regulation complexity 84, and those involved in cellular signaling 85 and protein-protein interaction networks 86. Functional properties such as transcriptional regulation, signaling and protein-protein interactions are associated with disordered proteins 12567897883. Our study highlights the association between predicted miRNA targets and protein disorder in a general way.

Mono-ubiquitination rather than polyubiquitination is the prevalent signal in intranuclear trafficking and triggers the first step of endocytosis 53878889. The transcripts relating to polyubiquitination GO categories are down-co-regulated with highly disordered proteins (Table 4). Our results show that expression levels for the transcripts encoding highly disordered proteins are anti-correlated with transcripts involved in proteolysis and ubiquitin-dependent cellular catabolism. The expression levels for the transcripts encoding highly disordered proteins are positively correlated with proteolytic inhibitors. Whilst post-translational modification of proteins by ubiquitin is a key regulatory event, de-ubiquitination proteases, the enzymes that remove and process ubiquitin from proteins, are known to be functionally important 90. Many of the de-ubiquitination proteases are cysteine proteases 90. Transcripts coding for cysteine protease inhibitors, endopeptidase inhibitors and protease inhibitors are up-co-regulated with highly disordered protein transcripts (Table 4), which may suggest that protease enzymes involved in de-ubiquitination are not expressed when the expression of disordered protein transcripts is high. We hypothesize that if a highly disordered protein has a ubiquitin tag, the function of this tag is more likely to be brought to fruition 8790. Highly disordered proteins with one or more predicted ubiquitination sites that are not constitutively or highly expressed may have a higher chance of being removed from the cell, as they are likely to have a higher density of ubiquitination sites. The (80,100]% disorder category is likely to have the highest density of ubiquitination sites (Figure 2e). Since ubiquitination is a feature of many biological processes 878889 the presence of these ubiquitin target sites may also be implicated in protein transport between membrane components, possibly serving as a sorting signal and/or a regulatory signal for internalization into the endocytic pathways.

Transcripts associated with GO terms involved with protein complex formation are co-expressed with transcripts of highly disordered proteins (Table 4). Transcripts belonging to the (60,80]% disorder class are co-expressed with transcripts involved in heterodimeric complex assembly whilst those that belong to the (80,100]% class are co-expressed with those transcripts involved in homodimeric complex assembly. Several disordered proteins are known to be involved in protein cellular complexes 57883, providing support for this finding. Transcriptome and interactome studies are known to provide complementary results 9192. The results indicating that chaperone binding correlates with highly disordered proteins (Table 4) are not conclusive and this is supported by a recent analysis 93; it is hypothesized that disordered proteins that bind chaperones do so to avoid aggregation and assist in complex assembly 94959697. Expression levels for transcripts involved in catabolic processes are down co-regulated with expression of transcripts encoding disordered proteins, that is expression levels of these transcript groups are simultaneously low (Table 4). This suggests that there is a reduction in the catabolism of biopolymers (such as proteins) and that the transcripts encoding highly disordered proteins and the resulting protein product remain in the cell longer to carry out function. The gene expression measurements were taken from 200 normal, 'resting' and un-stimulated tissues and this could partly explain the down-regulation of catabolic processes as well as other observations.

Human protein sequences were obtained from the Ensembl FTP website (Assembly version 35). Disorder predictions were carried out using DISOPRED2 6 with a 2% false positive rate. Our definition of a disordered protein states that it must contain at least one region of 30 contiguous disordered residues. This cut-off was based on previous work 98. Two or more disordered regions separated by a number of ordered residues were considered distinct disordered regions. Using MEMSAT3 99 and PFILT 100, we filtered our datasets for transmembrane and coiled-coil regions. This ensured a low false positive rate in our disorder predictions. An ordered protein is defined as one that has no disordered regions; that is, it does not contain a contiguous region comprising 30 or more predicted disordered residues. Using these definitions 98, proteins were classed as being ordered or disordered (Table 2).

A second classification scheme was devised to generate protein datasets reflective of high order and high disorder. To achieve this, we divided the proteome into 3 groups: those containing less than 10 residues of disorder; those containing 60% or more disordered residues; and those not fitting within either group. For the highly ordered protein dataset (Table 2), we chose a lower end cut-off of less than ten residues of disorder to allow for false positive disordered residues. For the highly disordered proteins (Table 2), an upper cut-off of 60% was used to ensure that this group contained genuine high levels of disorder in proteins. A third scheme to classify disorder in proteins was based on the percentage of disordered residues in the protein. The percentage of disordered residues in the protein datasets were binned (Table 2). The first bin [0,20]% has from 0% (inclusive) to 20% (inclusive) residues disordered; the second bin (20,40]% has >20% and ≤40% residues disordered and so on until the final bin (80,100]%, which has >80% and ≤100% residues disordered.

In previous work, Yang et al. 33 measured the mRNA decay rates of 5,245 human transcripts. The mapping of the EMBL (GenBank/DBJ) identifiers for these mRNA transcripts 33 to the Ensembl protein identifiers of our protein datasets was facilitated by the use of Biomart 106. Using their EMBL (GenBank/DBJ) identifiers, we mapped the decay rates to the Ensembl protein identifiers of our dataset of human proteins. This gave a dataset of 3,839 proteins, each of which has an associated experimentally determined mRNA decay rate. We separated our 3,839 protein dataset into 3 groups; highly disordered (74 proteins); highly ordered (536 proteins); and the remainder (3,229 proteins).

Recent work by Yen et al. 75 reported half-life protein stability measures for more than 8,000 human proteins using their global protein stability assay. This study is one of the most comprehensive studies of in vivo stability measured for proteins in complex cellular mixtures. We mapped these stability measures to our disorder protein dataset, which resulted in a set of 6,886 proteins. We separated this dataset into three groups, as described above. This resulted in 179 highly disordered proteins, 1,396 highly ordered proteins and 5,311 remaining proteins.

The predicted mRNA targets of mammalian miRNAs were downloaded from the TargetScanS website 44. The dataset downloaded (release 4.2; April 2008) contains 218,298 mRNA target predictions. The predictions were performed using TargetScanS 44. From this dataset, 7,928 unique genes (HUGO identifiers provided) were predicted to be targets of miRNA. The HUGO identifiers 107 for the mRNA genes were extracted into a gene list. An analysis pipeline was developed to establish whether a correlation exists between the disordered proteins and the transcripts targeted by miRNA. The first part uses the Ensembl database and the PERL API 108 to map the HUGO identifiers to Ensembl identifiers (the gene, the transcript and the protein) and extract the associated translated protein sequences. We identified 15,954 transcripts for the 7,928 HUGO gene identifiers. The second part categorizes the protein datasets based on the amounts of disorder (Table 2). With each protein category, a protein list is derived and protein sequence dataset is created. The third part compares each protein dataset with the translated products of the mRNA targets regulated by miRNAs by tallying identical protein sequences derived from the 15,954 transcripts and the transcripts from each disorder category. For a given dataset (Table 2), to calculate the percentage of protein transcripts that are predicted miRNA targets, the number of matches between the two datasets was divided by the total number of proteins in the disordered protein category under consideration and this fraction was multiplied by 100. Fisher's exact tests were carried out in R 109 to identify groups of transcripts that were enriched in miRNA target sites.

Ubiquitin targeting signal sites were predicted for the peptides encoded by the Ensembl transcripts using a neural network-based predictor. Our predictor uses a single hidden-layer back-propagation network trained to recognize features of ubiquitin targeting signals over a sliding window of 21 amino acids in the target sequence. The network was trained using a balanced dataset of Swiss-Prot annotated ubiquitination sites 110 and rigorously cross-validated using a jack-knife leave-one-out approach. The performance of our predictor on both the Swiss-Prot training data and a non-redundant experimentally defined dataset 110 was also determined (Figure S7 in Additional data file1). Predictions were only reported at a false discovery rate of <5% estimated from our model. Although no ubiquitination predictors were available when we carried out this study, the Ubi-pred tool 111 has subsequently been published. This tool obtains a final area under the curve (AUC) 112 of 0.85, sensitivity of 70.86, specificity of 0.954 and a Matthews correlation coefficient of 0.69 compared to our own ubiquitination predictor, which obtains a final AUC of 0.88, sensitivity of 83.44, specificity of 85.43 and a Matthews correlation coefficient of 0.69. These statistics provided the justification for using our own ubiquitination prediction tool over Ubi-pred 111 due to the observed lower false positive rate. The percentage accuracy is defined here as 100% - Percentage of errors and the AUC is the area under the receiver operating characteristic (ROC) curve (a standard score for prediction algorithms between 0 and 1).

As an independent test of our ubiquitin site prediction algorithm, we scanned a dataset of putative and experimentally determined ligase target sequences 113 for the presence of ubiquitin modification sites (Table S4 in Additional data file 2). In total, 72.6% of sequences (257 of 345) were predicted to contain at least 1 modification site. This demonstrates that our in silico predictions are in good agreement with independent experimental outcomes.