Sequencing of the human genome and collections of EST have facilitated global detection of alternatively spliced variants (5, 6). Because ESTs are generally derived from mature spliced mRNA populations, they provide a broad sample of mRNA diversity. Computational analyses of cDNA and EST sequences have suggested that alternatively spliced transcripts are produced from >50% of mammalian genes (7-9). Most alternative splicing occurs within protein-coding regions (10, 11). It has been suggested that alternative splicing is a general mechanism to increase the coding capacity and diversity of the genome in metazoans (12).

To measure genome-wide trends of alternative splicing, we performed a separate computational analysis that characterized the distribution of splice variants among different classes of genes. For these analyses, the reference set of 8,100 full-length cDNAs representing known human genes was extracted from SwissProt (release 41), a manually annotated database (23). The EST tissue source information was extracted from the TissueInfo database (24). Detailed descriptions and complete results of the computational analyses are included in Supporting Materials and Methods and Data Sets 1-6, which are published as supporting information on the PNAS web site; additional data also are available from the authors upon request.

Using blast, we compared the identified splice variants with sequences in the NCBI EST database that includes all of the publicly available ESTs. We found that ≈30% of the splice variants identified in the stem cell data sets were not found in ESTs from other tissues. We chose a set of 15 genes encoding diverse proteins, such as transcription factors and transmembrane receptors, for experimental confirmation. Prior knowledge of alternative splicing was not used in the selection of the gene sets. Using cloning and sequencing, we were able to confirm that 12 of 15 (80%) of these genes were expressed with the splice variations (Fig. 2 and Fig. 7, which is published as supporting information on the PNAS web site). Lack of confirmation for 20% of the selected genes was likely due to differences in the stem cell lines, isolation protocols, and cell culture conditions. The RT-PCR profiles across ES, HSC, and eight adult tissues showed that selected genes are expressed in stem cells with various degrees of specificity. Literature searches indicate that some of the confirmed genes play important regulatory roles in stem cell biology. For instance, it was shown that the Polycomb group transcription factor EZH2 is required for maintenance of undifferentiated cells during the blastocyst stage of early mouse development (27). The Bric-a-Brac/Poxvirus and zinc-finger domain protein Zbtb20 is a less characterized transcription factor; however, it has been shown to be involved in developmental neurogenesis (28). In the group of transmembrane genes, we identified splice variants for platelet/endothelial cell adhesion molecule-1, a transmembrane protein involved in migration of hematopoietic stem cells (29). Splice variations were also confirmed for the weakly characterized putative adhesion receptor FAD104 and the apoptotic TNF receptor TNFR-7 (Fig. 7). The majority of our experimentally confirmed variants are not described in the published literature (Table 3, which is published as supporting information on the PNAS web site).

To estimate a possible functional impact of splice variations in stem cells, the distribution of the identified variants was analyzed across signaling pathways found in the Proteomic Pathway Project (BioCarta, San Diego). Overall, the identified splice variants showed a statistically significant enrichment for genes encoding components of signaling pathways (P < 0.01), as assessed by the hypergeometric distribution (Supporting Materials and Methods). Because the available EST stem cell data are limited, it is difficult to determine whether some pathways are affected more than others. However, we found that alternative splicing modifies components of several signaling pathways that are known to govern stem cell self-renewal and differentiation (Table 4, which is published as supporting information on the PNAS web site). For instance, splice variations were detected in genes encoding components of the mitogen-activated protein kinase signaling pathway, such as Map4k2 and Mnk1 kinases. We also identified alternative splice variants for multiple genes involved in cell cycle and apoptosis. These included cyclin-dependent kinase 4, cyclin-dependent kinase inhibitor 1, cyclin H, caspase 11, and apoptosis regulator BID. The majority of the identified variants were not described previously. Complete results of the pathway analysis are available from the authors upon request. Collectively, our findings indicate that alternative splicing plays a major role in the diversification of numerous regulatory gene products expressed in stem cells.

Correlations Between Transcription and Alternative Splicing. To further explore the observation that alternative splicing extensively modifies regulatory genes, we analyzed the distribution of splice variants across various expression levels. Because the available stem cell EST collection is limited, we used human EST data from all tissues. More than 4,000,000 human ESTs were aligned to a nonredundant reference set of 8,100 full-length cDNAs derived from SwissProt. This database is manually annotated and therefore reliably defines bona fide gene products. Computational analyses identified 2,471 alternative splice sites within coding regions of human genes. The complete data set is presented in Data Sets 1-3.

To describe the frequency of alternative splicing in a quantitative manner, we defined a new parameter, an exon exclusion fraction, f_ex (Fig. 3). Computationally, the exon exclusion fraction is defined as the number of ESTs with an excluded exon divided by the total number of ESTs spanning a site of alternative splicing: where N_ex is the number of ESTs with an excluded exon and N_in is the number of ESTs with an included exon. Experimentally, the exon exclusion fraction is defined as the intensity of DNA bands representing the exon exclusion isoform divided by the intensity of all of the bands: where I_ex is the intensity of the exon exclusion isoform (lower band), I_in is the intensity of the exon inclusion isoform (upper band) and L is a fragment length factor. The fragment length factor, defined as the ratio between the long and short isoform lengths, is needed to correct the effect of DNA fragment length on UV absorption. Because tissue-specific EST libraries are not yet sufficiently comprehensive, our calculations were performed across all tissues. For example, 12 ESTs and 14 ESTs represent cases of exon inclusion and exon exclusion in the cAMP-response element-binding protein gene, respectively. In this case, the exon exclusion fraction is 14/(12 + 14) = 0.54.

To test for correlations between the levels of transcription and alternative splicing, the exon exclusion fractions at each splice site were analyzed as a function of EST numbers covering this site. As shown in Fig. 4A, genes represented by a high number of ESTs show a low frequency of alternative splicing; that is, the majority is represented by a single isoform (i.e., exon exclusion fractions are close to 0 or 1). We chose sets of genes represented with various numbers of ESTs to confirm the computationally identified trends (Fig. 4 B and C). Prior knowledge of alternative splicing was not used in the selection of the gene set. The selected genes with a high number of ESTs encode proteosome components PSB1 and PSB3, the exosomal component RR46, and glucosyl transferase EXT2. According to computational results, the exon exclusion fractions were close to 0.01 for these genes. When tested experimentally, some of these genes do not show detectable splicing variants, whereas others show a very low frequency of alternative splicing (Fig. 4B and Fig. 8, which is published as supporting information on the PNAS web site). In contrast, a higher frequency of alternative splicing was observed computationally and confirmed experimentally for genes represented with low number of ESTs (Fig. 4C and Fig. 9, which is published as supporting information on the PNAS web site). Selected examples include the ion channel P2X5, the hematopoietic transmembrane receptors IL-7R and CD3D, and transcription factor subunit CBFB. For these genes, the exon exclusion fraction was closer to 0.5.

Because ESTs generally originate from random sequencing efforts, the number of ESTs corresponding to a specific gene is representative of its expression level. Because only 5% of the ESTs in our data set were derived from the most contributing tissue (brain), we assumed that high EST representation characterizes ubiquitously expressed genes. We also assumed that a low number of ESTs characterizes tissue-specific genes. Our RT-PCR data support these assumptions (Figs. 4 B and C, 8, and 9).

To further confirm a correlation between a high frequency of alternative splicing and tissue-specific gene expression, we computationally compared distribution of exon exclusion fractions between tissue-specific and ubiquitous genes (Supporting Materials and Methods). Because EST libraries are not comprehensive, we normalized expression levels for genes that are represented with 10 or more ESTs across 102 tissues. Genes were defined tissue-specific if 30% or more of their ESTs originated from a single tissue (tissue specificity P value of <0.001). Genes were defined as ubiquitous if <20% of their ESTs originated from a single tissue (tissue specificity P value of >0.01). We found that the tissue-specific subset included 4-fold more genes with exon exclusion fractions within 0.5 ± 0.3, as compared with the subset of the ubiquitous genes. These experimental and computational data indicated that frequency of alternative splicing is higher for tissue-specific rather than ubiquitous genes. However, as shown in Fig. 4A, there are clearly examples of ubiquitously expressed genes with ratios close to 0.5.

Low Frequency of Exon Exclusion. Upon examination of the genes by RT-PCR, we noticed that they were usually represented by a single isoform. The formation of other isoforms is usually rare and tissue-specific. To explore this trend, we analyzed the distribution of the exon exclusion and exon inclusion events at identified alternative splice sites. If exon exclusion and inclusion were occurring at a similar frequency, we would expect to obtain a Gaussian-like distribution with a peak centered at 0.5. However, as shown in Fig. 5A, the inclusion or exclusion frequencies of individual exons were distributed in a skewed, highly nonrandom fashion. The exon exclusion fraction was <0.2 in nearly 60% of the analyzed alternative splice sites. Thus, at most sites of alternative splicing, exon exclusion was a rare event. Interestingly, very few sites showed an equal representation of both alternatively spliced isoforms. Four examples are presented in Fig. 5B, and the complete data set of 25 genes is included in Fig. 10, which is published as supporting information on the PNAS web site. As expected, the majority of the experimentally tested genes showed predominant formation of the long isoforms, whereas exon exclusion was rare and tissue-specific. The apoptosis regulator MCL1 and transcription factor TF3A gene products shown in Fig. 5B represent examples of predominant exon inclusion. In comparison, the chromodomain-helicase-DNA-binding protein-1 gene CHD1 is an example of predominant exon exclusion. The gene encoding amyloid-like protein APP2 was an example of approximately equal exon inclusion and exclusion frequencies. In general, the computationally determined exon exclusion frequencies were in good agreement with the experimental data (Fig. 5C). These findings suggest that, at the majority of alternative splicing sites, exon inclusion is constitutive and exon exclusion is a rare and tissue-specific event.

To deepen our understanding of alternative splicing and its regulation and functional consequences, we subsequently analyzed its trends across all tissues. We found that ubiquitously expressed genes show a very low frequency of alternative splicing. It may be that the low-frequency splice variants represent the occasional infidelity of the splicing machinery. Theoretical arguments have estimated a 0.001 frequency of such errors (33). This value is similar to the frequencies at which we detect splicing variants in genes that encode proteosomal components. In contrast, tissue-specific genes appear to show a high frequency of alternative splicing. Previous studies of individual genes have shown that splicing is coupled to transcription by protein-protein interactions between components of the transcription and splicing complexes (34, 35). Taken together, these results suggest that, on the genome-wide level, coupling of transcription and splicing results in diversification of tissue-specific and regulatory gene products, with little effect on ubiquitous “housekeeping” genes. A supporting evolutionary argument is that ubiquitous transcripts responsible for crucial and general cellular processes have evolved not to be modified, whereas diversification is advantageous for tissue-specific gene products. This explanation is further strengthened by our observation of fast evolutionary changes in alternative splicing patterns. We found that these patterns were conserved for only 20% of the examined orthologous genes in the human and mouse species, despite the general conservation of their exon-intron boundaries. These observations are in agreement with results of a recently published study (36) and consistent with previous conclusions regarding the rapid evolution of alternatively spliced exons (37). Lack of conservation of the alternative splicing patterns may contribute to the previously observed differences in functional properties of analogous cell types, such as mouse and human ES cells (38).

At the molecular level, alternative splicing results from blocking of constitutive splicing sites or activation of weak (cryptic) sites (2, 4). For a large set of alternatively spliced genes, we observed that exon inclusion is predominant, whereas exon exclusion is rare and often tissue-specific. A similar conclusion was obtained in previous computational studies (37). Our experiments confirmed these computationally derived trends, which implies that exon inclusion is a default option in the overall expression process of these genes. Based on these observations, we hypothesize that repression of constitutively used splice sites in primary transcripts is responsible for the formation of most splice variants. Furthermore, such blocking is likely to occur in a tissue-specific manner. However, the observed bias toward exon inclusion may partially reflect an artificial effect from accumulations of ESTs with rare splice errors in ubiquitous genes, as discussed above.