To investigate the functional properties of human 5UIs, we used NCBI's Reference Sequence (RefSeq) collection. These are curated, full-length sequences with annotated UTR boundaries, and expression data are available for many of them. The lack of a translation reading frame makes the computational prediction of splice sites in 5'UTRs inherently more difficult 37, necessitating the choice of such a validated set. In humans, approximately 8.5k (35%) out of 24.5k RefSeq mRNAs contained at least one intron in their 5'UTR (Additional file 1). Previous estimates of the percentage of genes with 5UIs ranged between 22% and 26% 18 and 38% 19 in humans, suggesting that the RefSeq collection had no major bias in terms of presence or absence of 5UIs compared to other previously used datasets. The distribution of total 5'UTR intronic length for genes in our dataset was also similar to that observed previously (Figure 1a). The inter-quartile range of total length of 5UIs within each gene was approximately 1.3 - 16 kb. Some 5UIs were extremely long -- 16% were longer than 27 kb, the length of the average protein coding gene in the human genome 38, and 5% were longer than 76 kb (Figure 1a). As previously reported 1819, most genes had few 5UIs. More than 90% had a single intron, and the percentage of genes with two or more introns decreased exponentially (Figure 1b).

We next considered the relationship between the total lengths of 5'UTR exons and of 5UIs. Even though there was a correlation between the lengths of 5UIs and 5'UTR exons overall, this correlation was slight and was driven by the genes with the longest 5UIs (Figure 1c; Pearson correlation coefficient or Pearson correlation coefficient (PCC) = 0.21, P < 2.2e-16). In fact, when genes with 5UI lengths in the lowest 25th percentile were analyzed, the correlation was no longer significant (Figure 1c; PCC = -0.005, P = 0.84). A statistically significant, albeit slight, correlation was found for genes with 5UI length below the median (Figure 1c; PCC = 0.07, P = 8.4e-05). Among the genes with 5UIs, a similar relationship was evident between the total length of 5UIs and the total length of the remaining introns (Figure 1d). Although these two variables were significantly correlated (Figure 1d; PCC = 0.18, P < 2.2e-16), the relationship was clearly driven by the genes with longer 5UIs. When genes with 5UI lengths either in the lowest 25th or 50th percentile were considered, correlation was negligible (Figure 1d; PCC = -0.02 and 0.04, P = 0.53 and 0.04, respectively).

Thus, genes with long 5UIs tend to have a high total intronic length and longer 5'UTR exons. While this tendency holds in genes with additional introns, several genes with total 5UI lengths greater than 10 kb lack any coding-region or 3'UTR introns (Figure 1d). On the other hand, amongst genes with short 5UIs, the total length of 5UIs is uncorrelated with the lengths of either 5'UTR exons or the remaining introns.

We next examined gene expression-related predictions of the two principal models of intron evolution. Previous studies have suggested that the genes with the highest expression levels are selected to have shorter introns 23. If a similar selective pressure were acting on 5UIs (in conjunction with neutral evolutionary processes 19), one would expect a tendency towards reduced gene expression level as a function of increased 5UI length in a subset of genes. We therefore compared gene expression from 79 tissues as a function of the total 5'UTR intronic length. We divided 5UI-containing genes into three categories with respect to the total 5'UTR intronic length (short, 0 to 25%; intermediate, 25 to 75%; long, 75 to 100% in length). The short 5UI-containing genes were highly overrepresented in the top 1% of mean expression level for the genes with 5UIs (Fisher's exact test, P = 3.3e-15) and also in the top 5% (Fisher's exact test, P = 1.7e-14) (Figure 2a). These genes were 12.7 times more likely than all other genes with 5UIs to be in the highest 1% of mean expression and 3 times more likely to be in the highest 5% of mean expression. There was also a global trend for genes with short 5UIs to be expressed at a higher level compared to genes with longer 5UIs (25 to 100 percentile in length; one-sided Wilcoxon rank sum test, P = 2.98e-05; Figure 2a).

The enrichment for high expression in genes with short 5UIs held even when genes with the longest 25% of 5UIs were removed. In this case, the genes with the highest 1% and 5% expression were, respectively, 9.5 times and 2.5 times more likely to have short 5UIs as opposed to intermediate length 5UIs (25 to 75 percentile in length; Fisher's exact test, P = 1.53e-11 and P = 3.21e-10, respectively).

The most highly expressed 5UI-bearing genes show a striking tendency to harbor short 5UIs. Of all 5UI-containing genes, 26% had a total 5UI length below 1.3 kb. By contrast, the corresponding fractions for genes in the top 5% and 1% by expression were 50% and 83%, respectively. We then separated short 5UI-containing genes into two groups: the most highly expressed genes (top 5% in expression); and the remaining genes. For the most highly expressed genes, the inter-quartile range of total 5UI length was 215 to 734 nucleotides compared with 289 to 870 nucleotides for the remaining genes (Figure 2b). Thus, the most highly expressed genes in humans are very strongly enriched for short 5UIs.

Interestingly, no expression dependence was observed among genes with intermediate or long 5UIs: genes with long 5UIs (top 25th percentile in length) did not tend to be expressed less than those with the intermediate length 5UIs (Wilcoxon rank sum test, P = 0.25). Also, no statistically significant depletion for the long 5UI category was observed in either the top 1% or the top 5% expression group (Fisher's exact test, P = 0.29, odds ratio = 0.25, and P = 0.017, odds ratio = 0.58, respectively). Thus, we did not observe the inverse relationship between expression and total 5UI length that might have been expected under the energetic cost model.

Next, we considered all RefSeq genes and asked whether having an intron in the 5'UTR has an effect on overall expression. We found no differences in 5UI representation in the top 1% or the top 5% of the mean expression groups. Furthermore, no difference was detected in the distribution of mean expression between genes with and without 5UIs (two-sided Wilcoxon rank sum test, P = 0.17). However, genes with short 5UIs were 1.8 times more likely to be in the top 5% and 3.3 times more likely to be in the top 1% in overall expression level than genes with no 5UIs (Fisher's Exact Test, P = 3.15e-08 and P = 7.57e-07, respectively) than genes with no 5UIs (Figure 2c). Thus, the presence of short 5UIs is correlated with high mean expression.

The observed expression trends could reflect the influence of genomic features other than 5UIs. Yet, short 5UIs do not seem to predict a short total length of either non-5'UTR introns or 5'UTR exons (Figure 1c, d). Furthermore, when genes in the top 5% in mean expression were divided into two groups with respect to 5UI presence or absence, we observed no differences in total non-5'UTR intron length between genes with 5UIs and those that lack these introns (Wilcoxon rank sum test, P = 0.20, data not shown). Therefore, the tendency of highly expressed genes to have short 5UIs is unlikely to be confounded by the effects of 5'UTR exons or the remaining introns.

For genes with the highest expression levels, these results are in contrast to the neutral model of 5UI evolution, which predicts that 5'UTR intronic length should not depend on expression level. These results are also not explained by the energetic cost hypothesis, which would predict that genes with the highest expression levels should be less likely to have 5UIs. In stark contrast to the predictions of each model, we found the most highly expressed genes to be significantly enriched in short 5UIs. Furthermore, the energetic cost hypothesis would also predict a linear decrease in the total 5UI length as a function of increasing gene expression. Yet, we found no overall differences with respect to 5UI length except for the most highly expressed genes. Even though a neutral model of 5UI evolution is plausible for most genes, our results for the most highly expressed genes are inconsistent with both neutral and energetic cost models (Figure 2d).

We next used expression to assess the applicability to 5UIs of the other major hypothesis of intron evolution, the 'genome design model', which predicts that intermediate or long introns should be enriched in tissue-specific genes as a consequence of complex regulation. As originally outlined, the genome design model explicitly disregards 5UIs 27; however, a direct corollary of this hypothesis is that genes with higher variance in expression across tissues should have intermediate or long introns in their 5'UTRs as well.

We sought to address two potential sources of bias. First, gene expression levels vary greatly and variance is strongly correlated with mean expression. Therefore, we calculated the standard deviation-to-mean ratio (coefficient of variation or CV) 39, a normalized measure of dispersion, for each gene across all tissues. Second, due to technological limitations of expression arrays, precise measurement of expression level is more difficult for genes with low or no expression in a given tissue; therefore, artificially high variance in expression might be observed for genes with low mean expression across all tissues. We therefore calculated a robust measure of dispersion that minimizes this effect:

where CV_x is the CV of expression of gene x across all tissues, y_x represents the vector of CV values for all 201 genes in a window centered around gene x, while μ_1/2 and MAD represent the median and median absolute deviation, respectively. As expected, genes with low expression tended to have much more variability across tissues (Figure 3a). Based on the observed trend line, the genes with the lowest 25% expression were removed from further analysis (Figure 3a). The remaining genes were sorted into three categories with respect to the total intronic 5'UTR length as before (short, 0 to 25%; intermediate, 25 to 75%; long, 75 to 100%). We found no significant differences between these groups with respect to inter-tissue variability as measured by the coefficient of variation (Figure 3b; Kruskal-Wallis rank sum test, df = 2, P = 0.23). We then examined the lengths of the introns as a function of variability in expression (Figure 3c). The genes with the highest 5% variability across tissues did not differ from the other genes with respect to their 5UI lengths (Wilcoxon rank sum test, P = 0.07, 95% confidence interval between -0.008 and 0.25), but the genes with highest 1% across-tissue variability tended to have slightly shorter 5UIs (Wilcoxon rank sum test, P = 0.006, 95% confidence interval between -0.67 and -0.11). Genes with short 5UIs were also overrepresented in the top 1% across-tissue variability category (Fisher's Exact Test, P = 0.005, odds-ratio = 2.7). Our results suggested that length of the 5UI was not a major factor in determining across-tissue variability but there was a preference for shorter 5UIs in the most variable genes.

Although our approach reliably captures across-tissue variability in gene expression, it disregards any potential effects of 5UI presence or length on how widely a gene is expressed. To consider the potential impact of such effects, we calculated the number of tissues in which expression was detected for each gene. Based on our analysis presented in Figure 3a, we defined a given gene as 'present' in a given tissue if its expression was greater than the 25th percentile in the distribution of mean expression over all tissues, calculated for all genes. Genes were placed into one of five classes according to the number of tissues in which they were present. No significant difference was detected amongst the corresponding five distributions of total 5UI length (Figure 3d; Kruskal-Wallis rank sum test, df = 4, P = 0.19). Furthermore, the distribution of number of tissues in which each gene was present did not differ between genes containing and lacking 5UIs (Figure 3e). These results clearly contradict predictions of the 'genome design' hypothesis, in that narrowly expressed genes did not show a greater tendency to contain 5UIs nor did they tend to have longer 5UIs. These results strongly suggest that the evolution of 5UIs is not driven primarily by the selective pressures proposed by the 'genome design' hypothesis.

Under the neutral model, genes with 5UIs should be uniformly distributed across functional groups. We used Gene Ontology (GO) function annotations to determine which groups of genes are enriched or depleted in 5UIs, if any. Two popular functional trend analysis tools, FuncAssociate 40 and GoStat 41, were used for this analysis. One key challenge was the translation of the gene identifiers from RefSeq RNA IDs to those used in the GO database. There are different approaches to this problem and the two software packages differ from each other in this respect. FuncAssociate uses the Synergizer 42 software to resolve the problem of synonyms while GoStat uses definitions in the UniGene database as well as the information provided in the GO databases. Both software packages yielded very similar results, suggesting that our general conclusions were independent of the methods of synonym resolution or enrichment calculation.

A significant overrepresentation of genes with 5UIs was found in many regulatory pathways (Table 1). Non-receptor protein tyrosine kinases (NRTKs) formed the most highly overrepresented group, followed by genes involved in the regulation of actin organization, transcriptional regulators, and zinc ion binding proteins (Table 1). NRTKs lack transmembrane domains and therefore do not recognize extracellular ligands, unlike the majority of protein tyrosine kinases. Nevertheless, they play crucial roles in nearly all aspects of biology and are implicated in many cancers (reviewed in 43). Among NRTKs, genes harboring 5UIs encode key regulatory kinases, such as the proto-oncogene tyrosine kinase SRC, c-src tyrosine kinase (CSK), janus kinases (JAK), spleen tyrosine kinase (SYK), tec protein tyrosine kinase (TEC), and Bruton agammaglobulinemia tyrosine kinase (BTK) among others.

To gain insight into the evolution of NRTK 5UIs, we identified orthologous genes in mouse and rat genomes corresponding to each human NRTK. We collected 5'UTR features for these genes in each genome using RefSeq annotations (Additional file 2). More widely studied organisms tend to have more accurate transcript structures and include many more splice variants in the RefSeq collection. For example, 18 human genes were represented by more than one transcript, while only four mouse and no rat NRTKs had more than one splice variant. The paucity of transcripts in some mammalian species is more likely to have arisen from limited testing rather than biology, given recent studies suggesting that alternative splicing is ubiquitous across several taxa 9.

UTRs are also generally less well defined in less intensively studied organisms. For example, ABL2, BTK, FRK and SRC all lack defined 5'UTR boundaries in the rat RefSeq collection, even though EST evidence suggests that SRC, BTK and ABL2 all have 5'UTR-containing transcripts (data not shown). Another current limitation is ambiguity in identifying the specific branch in which a given deletion or insertion event took place. Despite these shortcomings, a comparison of orthologs already provides insight into the dynamics of the evolution of 5UIs in NRTK genes.

When every ortholog of a given NRTK had at least one annotated 5UI, the lengths of those introns were generally highly correlated (Figure 4a). Given the number of different splice variants for each human gene, we used three different approaches to calculate the 5UI length for each gene. We either used the mean length of splice variants with non-zero 5UI lengths, or picked the variant with the longest 5UIs, or the one whose length was closest to its ortholog in either of the rat or mouse genomes. All three measures resulted in high correlation overall between 5UI lengths across species (PCC ranged between 89 and 91% for human-mouse and 79 and 89% for human-rat comparisons; P < 0.0001 for all; Figure 4a). As expected from evolutionary distances, the highest correlation in 5UI lengths was observed between rat and mouse orthologs of NRTKs (PCC = 93%, P = 1.4e-07).

Despite a generally strong correlation in 5UI length among orthologs, some sets of orthologs had a widespread distribution of length changes. While the total 5UI length of FES changed by less than five nucleotides in all possible comparisons, rat PTK2 and mouse PTK2 5UIs differed by approximately 63.5 kb (Figure 4b, c). The length conservation observed for the FES 5UI is notably consistent with the high regulatory potential previously calculated for this 5UI 44 (Figure 4d). More broadly, introns containing regulatory regions might be expected to have high length conservation.

When each orthologous group of NRTKs was analyzed, we found variability with respect to presence/absence of 5UIs in some of these groups. For example, STYK1 and WEE1 both had 5UIs in humans, but not in mouse or rat (Figure 4b, c). In the case of human WEE1, two transcripts were identified in the human RefSeq collection - while one variant had a 512-nucleotide 5UI, the other variant lacked 5UIs entirely. This observation suggested the possibility that intron-containing variants might be present in mouse and rat without being represented in the RefSeq transcript collection. Indeed, we found EST evidence that rat WEE1 has a splice variant that includes a 5UI [GenBank:CK603528.1]. On the other hand, mouse FRK (Figure 4b) and rat TXK (Figure 4c) had 5UIs while their orthologs did not. We also observed several NRTKs having 5UIs in two of the species but not in the other one. For example, both human and mouse orthologs of LCK, BTK, CSK, TNK1, and YES1 had annotated 5UIs, while both human and rat orthologs of JAK3 and TEC had annotated 5UIs (Figure 4b, c). Our results suggest that NRTK 5UIs are frequently conserved, a conclusion that would be further strengthened should the apparent gain/loss events be attributable to incomplete transcript annotation.

The appearance of 5UIs in most human NRTKs (Table 1) suggested the potential for a common regulatory mechanism acting via shared motifs. To search for shared and conserved motifs in these introns, human NRTK 5UI sequences were located in human-to-mouse and human-to-rat genome alignments. For 37 out of 42 human NRTKs, more than 10% of the 5UIs could be aligned to both genomes; only these conserved fragments were used for motif finding. Overrepresented RNA and DNA motifs were sought in these aligned sequences using the PhyloGibbs software 45. In our search for overrepresented RNA elements, we identified two complementary motifs, so that the motif in these 5UIs is more likely to be relevant at the DNA level. A representative DNA motif (Figure 5a) with the highest log-posterior-probability was compared to the TRANSFAC v11.3 database of known transcription factor binding sites and to a list of conserved human predicted motifs 46 using the STAMP website 47 (Figure 5b, c). In both comparisons, the known binding site motif of the MAZ transcription factor was the most likely match. However, this does not rule out the possibility of this motif being the target of another DNA binding protein.

5UIs are, by definition, the most 5'-proximal introns in their transcript. However, not all 5'-proximal introns need lie within the 5'UTR. We sought to understand whether the observed functional properties of 5UIs were shared with 5'-proximal coding region introns (5PCIs). Given that the median position of the first 5UI was approximately 130 nucleotides away from the transcription start site regardless of the number of 5UIs 19, we defined the genes without a 5UI but with a coding region intron within 150 nucleotides of the transcription start site as 5PCI-containing genes. This criterion resulted in 24% of 5UI-lacking genes having a coding region intron that was deemed to be a 5PCI.

We next used GO annotations to compare the functional properties of 5UI-lacking genes with 5PCIs to those without 5PCIs. We observed the strongest enrichment of 5PCIs among genes in the following functional groups: MHC protein complex 1, cytosolic ribosome, hemoglobin complex, glutathione transferase activity, and transmembrane transporters (Additional file 3). This result contrasts the observed enrichment of 5UIs in regulatory genes. The differences in the enrichment profiles suggest that distinct functional groups of genes prefer early introns in either the 5'UTR or the coding region but not in both.

To assess the possible effect of 5' proximity on gene expression, we analyzed microarray data from the human gene expression atlas for 5UI-lacking genes. We found that genes with 5PCIs were more highly expressed on average (one-sided Wilcoxon rank sum test, P = 6e-08; Figure 6). We also observed a 2.3- and 3.7-fold enrichment for genes with 5PCIs among the most highly expressed top 5% and 1% of genes, respectively (Fisher's Exact Test, P = 4e-15 and P = 4e-09, respectively; Figure 6). The correlation between high expression and 5PCI presence was evident without any consideration of these introns' lengths. In contrast, no expression difference was observed between genes with or without 5UIs, on average, but short 5UIs were highly enriched among the most highly expressed genes (Figure 2c). These results suggest that early introns (both 5PCIs and 5UIs) are associated with the most highly expressed genes, but that this correlation is limited to short introns for 5UIs.

NCBI's human Reference Gene Collection (RefSeq) 63 and the associated annotation table were downloaded from the UCSC genome browser 64, genome assembly of May 2004. The annotation table was parsed using the Galaxy website 65 (as of June 2007) to obtain 5UI coordinates. Specifically, we extracted all introns annotated to lie between two 5'UTR exons. Then we removed all the cases where another splice variant was present in the RefSeq collection such that any sequence within the intron was part of the coding region. Hence, all the introns in our final dataset were strictly present in the 5'UTR according to the annotation of RefSeq genes. 5'UTR exon coordinates were similarly retrieved as of June 2007. Recent studies suggest that nearly all human genes are alternatively spliced 131415. However, it is not clear what fraction of these events have biological significance as opposed to reflecting random noise associated with the less than perfect fidelity of the splicing machinery. Only when multiple independent sources of evidence support tissue-dependent alternative splicing can we be confident that these variants have real biological significance. Therefore, we used RefSeq transcripts, which are (unlike ESTs) manually curated and supported by multiple sources of evidence. For the comparisons between total lengths of 5UIs and the rest of the introns, we extracted coordinates of all non-5'UTR introns from the RefSeq annotation table (as of May 2009). A complete list of the genomic coordinates of 5UIs examined in this study is available as Additional file 1.

The microarray data were downloaded from Gene Expression Atlas, which included expression data from 79 different tissues in humans 66. We used the gcRMA-normalized data from the Affymetrix U133a and GNF1H arrays. Synergizer 42 was used to associate RefSeq genes with probe sets on the U133a array and custom Perl v5.8.8 scripts were used to parse the GNF1H annotation table (available on the Gene Expression Atlas website). The resulting correspondences of RefSeq IDs to probe sets on the GNF1H and U133a microarrays were merged to obtain a final mapping. Where multiple probe sets corresponded to a single RefSeq ID, the arithmetic mean of the expression values of all the probes was used to obtain a representative expression level for that RefSeq ID in each tissue. A single region of the genome can correspond to more than one RefSeq ID due to alternative splice variants and/or alternative promoters, and there were cases of a single probe set corresponding to multiple RefSeq IDs. To avoid overweighting such regions, we removed RefSeq IDs such that there were no duplicates. The representative RefSeq ID from each such probe set was chosen uniformly at random. For each gene with a 5UI, we calculated the mean expression level across all tissues and divided the genes into three groups with respect to total 5'UTR intronic length: short, 0 to 25%; intermediate, 25 to 75%; long, 75 to 100% in length. All expression analysis was performed using the R software package v2.6.0. In addition, the 'hexbin' 67 and 'zoo' 68 packages for the R platform were used.

GoSTAT 41 and FuncAssociate 40 were used for functional trend analysis. We restricted the space of genes to all genes in the RefSeq collection because we used annotations in this collection to determine the set of genes with 5UIs. We used the RefSeq IDs as input for analysis with both programs. FuncAssociate uses Synergizer 42 to resolve the synonyms using Ensembl as the authority. To quantify the effect size, all the statistically significant GO categories that are enriched in the genes with introns are sorted according to their log₁₀ odds ratio. All reported log odds ratios were obtained from FuncAssociate. Similar results were obtained using GoSTAT (data not shown).

To study the evolution of 5UI presence and length among NRTKs, we first identified orthologs of human NRTKs in the mouse and rat genomes. We used NCBI's Homologene Release 64 69 (as of September 2009) to identify 'true' orthologous genes. Based on a recent evaluation of different approaches, Homologene showed greater specificity than other comparable orthology sources for the purposes of detailed phylogenetic and functional analysis 70. We extracted the corresponding RefSeq IDs for each of the human NRTKs, and their mouse and rat orthologs. Then, we downloaded the RefSeq annotation tables for current genome builds (hg19, mm9, and rn4; as of September 2009) and used these annotations to determine 5UI lengths. All statistical analyses were performed using R software package v2.6.0. The raw data used in this analysis of human NRTKs are provided in Additional file 2.

The coordinates for the non-receptor tyrosine kinase genes that harbor introns were converted to human genome build hg18 using the LiftOver utility tool obtained from the UCSC Genome Browser website 71. If there were known alternative splice variants in the RefSeq database, the longest intron was used for motif discovery purposes. Multiple alignment blocks for the human, mouse, and rat genomes (builds hg18, mm8, and rn4, respectively) were extracted from the 17-way multiZ alignment at the UCSC Genome Browser. These alignment blocks were merged using the Stitch MAF blocks utility on the Galaxy website 65 to obtain a final alignment of the human non-receptor tyrosine kinases to the mouse and rat orthologs. We obtained alignments that covered more than 10% of the length of the 5UIs for 37 human NRTKs, and excluded the other five introns from the subsequent motif discovery steps.

PhyloGibbs v1.2 was used in motif finding 4572. Different phylogenetic trees were tested but they did not significantly affect the results (not shown); therefore, all the results we report here were generated using the (hg18:0.5,(mm8:0.8, rn4:0.9):0.6) phylogeny specified in Newick tree format. Both RNA and DNA motifs (that is, forward strand only and both strands, respectively) were searched and the intronic sequences were used to define the background nucleotide distribution of the region to account for differences in nucleotide composition of 5UIs. The resulting motifs were represented by position-specific scoring matrices. The STAMP 47 web site was used to find similar motifs in the TRANSFAC v11.3 database as well as in a comparative genomics study in humans 46. Default parameters were used in all comparisons.

To determine the lengths and positions of various genomic features, we first compiled a list of all RefSeq IDs. A single ID can correspond to multiple transcripts either that are expressed from the same or different regions in the genome. Such IDs can be associated with different transcript structures, and are therefore removed from further analysis. RefSeq IDs corresponding to genes in the hypervariable hla-locus were similarly represented multiple times in the RefSeq collection. In these cases, only the version in the reference genome was retained for further analysis.

After these initial filters, we calculated total lengths of 5UIs, 5'UTR exons, and other introns for each remaining RefSeq transcript. The position of the first coding intron was determined using the coordinates of all introns from the RefSeq annotation table that was retrieved as of May 2009. There were multiple identifiers for different splice variants that were transcribed from the same genomic location in the RefSeq collection. To avoid any systematic biases, we compared three different approaches in selecting RefSeq transcripts for further analysis. First, we kept all transcripts regardless of how many were transcribed from a given loci. Second, we determined equivalence classes of RefSeq transcripts, such that two IDs were in the same set if their transcription intervals (from start to stop position) overlapped by more than 20 base pairs. Then, we randomly removed RefSeqs transcripts such that only a single representative transcript remained for each equivalence class. Third, exact duplicates with respect to the 5'UTR were removed. Specifically, if two or more RefSeq IDs had the exact same 5'UTR, a single identifier was selected as a representative for that particular region. Splice variants that differ in their 5'UTR were not removed because these provide additional information about the lengths of 5'UTR introns and exons. All three methods yielded similar results and led to identical conclusions. Therefore, only one representative method is shown in the figures. The third method conveys the most information when discussing total 5UI lengths and hence was used in Figure 1a. By contrast, considering one representative from each transcriptional unit is more relevant when analyzing the correlation between two genomic features. Hence, the second method was used for Figures 1c, d.

For the specific GO categories used in our analysis, all the genes in a given category were retrieved from the human GOA database 73. The corresponding RefSeq identifiers were determined using the Synergizer software 42. Total exonic length and intronic length were calculated for all these genes as described above.