Identification of tumor-associated cassette exons in human cancer through EST-based computational prediction and experimental validation

Alternative splicing (AS) is a pivotal regulation mechanism allowing the expansion of the genome expression potential through the generation of multiple transcripts from a single gene. Indeed, over 90% of multi-exon genes undergo AS [1] generating on average a tenfold expansion of the transcriptome complexity [2]. Alternatively spliced exons have splice sites that can be specifically recognized depending on tissue specificity, developmental stage, external stimuli, cellular stress, or even pathological conditions [3]. So the presence or the expression level of specific splice variants can be indicative of a pathological or physiological condition or even be the cause of a disease. Recent studies demonstrated that alternative isoforms can be linked to many pathologies among which cancer, and play important roles in their etiopathogenesis [4]. Indeed, changes in the splicing pattern of a gene may affect different steps in the life of a cell (e.g. cell growth, adhesion, migration, invasion and cell death) eventually leading to subsequent tumor formation [5]. Several evidences of the association between alternative splicing and cancer have been reported. For example, it has been shown that about half of all active alternative splicing events in ovarian and breast tissues are altered in tumors, resulting in a massive tissue type-specific reprogramming of alternative splicing [6]. These alterations seems to be linked to the RNA binding protein FOX2 whose expression is downregulated in ovarian cancer and whose splicing results extensively changed in breast cancer samples. More recently, it has been reported that the presence of an alternatively spliced isoform of the receptor for factor VII/VIIa, named Tissue Factor (TF), is a prognostic marker in patients with non-small cell lung cancer [7]. Indeed, this soluble isoform of TF lacking exon 5 that encodes the transmembrane domain has been detected in various cancerous tissues and several studies suggested that its expression promotes tumor growth and is associated with increased tumor cell proliferation and angiogenesis in vivo [8]. Association of splicing with diseases does not necessarily imply that alternative splicing is a direct cause of cancer: aberrant splicing may be consequence of indirect effects produced by the disease-induced stress or may be caused by the general misregulation of cellular functions [9]. Nonetheless, the identification of a robust association between diseases and specific splicing patterns could be very useful to define signatures that can be used as diagnostic or prognostic indicators, allowing to obtain medically relevant information such as tumor stage and profiles [10, 11].

The availability of an increasing number of genome and transcript sequences, including Expressed Sequence Tag sequences (ESTs) and more recently RNA-Seq data, provides an essential information source for the computational detection of the alternative splicing pattern of genes [1, 2, 12‐15].

Currently, dbEST comprises more than 62 million ESTs from several hundred organisms, the highest numbers for individual organisms being over eight million for human [16]. Several bioinformatic tools have been developed for genome-wide detection of alternative splicing based on pairwise alignments of cDNA and genomic sequences [17, 18]. To improve the prediction accuracy, we developed a new tool, we named ASPic. ASPic implements an optimization strategy that performing a multiple alignment of all available transcript data (including full-length cDNA and EST sequences) to the genomic sequence detects the intron set that minimizes the number of splicing sites. It also generates, through a DAG-based combinatorial procedure, the minimal set of non-mergeable transcript isoforms compatible with the detected splicing events [19]. The reliability of splicing isoforms detected by ASPic has been recently established through a comparative assessment [20].

EST sequences have linked information about their source sample (commonly including tissue, cancer, gender, and developmental stage as well as organism information) [21]. In particular, 45% of EST data are collected from cancer tissues thus providing an invaluable source of information to study association of alternative splicing with cancer. Furthermore, the longer read length of EST data with respect to those obtainable from current next generation sequencing platforms may provide relevant information on exon connectivity, thus allowing a more accurate reconstruction of full length alternative transcripts. Assuming that the expression level of a splice variant in a certain source sample is roughly proportional to the number of related ESTs in the cDNA library [16], we may provide a qualitative and quantitative association of the detected splice sites according to tissue-specificity and tumor status.

Recently, several studies reported the identification of cancer specific variants by combining the computational analysis of the EST data with experimental validations [22‐24].

The availability of the unprecedented repertoire of 319,092 alternative splicing isoforms from 18,174 genes collected in ASPicDB [2] predicted by a comprehensive analysis of EST data included in the Unigene database [25] prompted us to develop a reliable strategy for the identification of cancer-associated isoforms. In particular, we focused our attention on the identification of alternative normal- and tumor-specific cassette exons which are particularly suitable as diagnostic or prognostic cancer biomarkers or as therapeutic targets. Remarkably, we newly discovered several cassette exons with a significant differential expression in normal and cancer conditions. For some of them, we carried out an experimental validation by RT-qPCR assays on commercial RNA from the same anatomical system where they were predicted. Considerably, we assessed the biomarker suitability of some cassette exons, found differentially expressed in the nervous system, in glioblastoma patients. Our results provide support to our computational analysis and suggest that the relative ratio between normal and tumor alternative transcripts may represent a promising tumor biomarker.

To carry out our genome-wide analysis of cancer related splicing isoforms, we considered all data collected in our ASPicDB database [2] including over 18,000 genes and 300,000 alternative transcripts.

As depicted in Figure 1A, the number of available ESTs was quite heterogeneous for the different tissues. More than 900,000 ESTs derive from central nervous system (CNS) and - on average - about 210,000 ESTs are available for all the other considered tissues. The ratio between tumoral and normal ESTs is close to one in tissues like gastrointestinal tract (INT), kidney (KID) and testis (TES), as the number of available ESTs in the two conditions is almost the same. Tissues like colon (COL) and prostate (PRO) are enriched by sequences obtained from tumoral libraries, conversely placenta (PLA) and central nervous system (CNS) contain more ESTs obtained from normal tissues.

The overall expression of each gene (represented by at least 10 ESTs) has been evaluated in all the considered anatomical groups and in both the normal and neoplastic status. For each gene - and for each tissue - we compared the number of ESTs from normal and neoplastic samples. Genes were grouped into five categories: genes exclusively expressed in normal condition (NN); genes exclusively expressed in neoplastic condition (TT); genes over-expressed in normal condition (N); genes over-expressed in neoplastic condition (T); genes with comparable expression in both normal and neoplastic conditions (E). Figure 1B summarizes, for each tissue, the distribution of genes in the five different classes. As expected, the number of classifiable genes in each tissue - although relatively small compared to the total number of analyzed genes - reflects the abundance of available ESTs. As a matter of fact, the highest number of classified genes (> 7000) belongs to CNS. On average, 40% of genes do not show a statistically significant differential expression in the two conditions (class E). The remaining 60% of genes are distributed among the 4 different classes (TT, T, NN, N). Bone marrow (BMA), breast (BRE), skin (DER), liver (LIV) and placenta (PLA) show a significant bias toward genes over-expressed in tumors (T and TT). Conversely, central nervous system (CNS) and prostate (PRO) are enriched in genes mainly expressed in normal tissues (N and NN). The genes exclusively expressed in the neoplastic condition in the different tissues, which could be considered tumoral markers in a specific tissue, are listed in Additional File 2.

In order to assess if genes mainly expressed in tumor cells with respect to genes resulting mainly or equally expressed in normal tissues were enriched in specific functional categories, we performed a statistic analysis using the Gene Ontology annotation. The occurrence of each GO term in cancer associated genes (T and TT classes) and in genes classified as N, NN or E was compared to its occurrence in the whole genes set thus identifying significantly over-represented ontologies by applying Fisher's exact test. Additional File 3 lists the top 10 most frequent terms in the two genes sets. Our results, not surprisingly, show that genes mainly expressed in normal tissues are related to general metabolic processes, while genes associated to cancer show - as expected - a sizable enrichment of processes related to cell proliferation, apoptosis and cell cycle.

In order to identify cancer or normal specific introns, we considered in our analysis all transcript-supported splices of genes belonging to the T, N and E classes, excluding genes belonging to the TT and NN classes. As shown in Figure 1C and detailed in Additional File 4, the majority of classified splicing events belongs to CNS with 18,436 classified introns, due to the higher number of ESTs available for CNS (Figure 1A). With few exceptions (e.g. colon, prostate and stomach), a general prevalence of tumor associated introns can be observed, with eye and placenta showing a noticeably 96% and 92% of tumor associated introns. This pattern is not related to the differential abundance of normal and tumor specific ESTs, as shown in Figure 1A. Indeed, for colon, prostate, and stomach, tumor associated-ESTs largely outnumber normal ones (Figure 1A).

By using the methodology described in Methods, we classified 26,974 splice sites as tumor- (T or TT) and 21,120 as normal- (N or NN) specific from 9513 genes. Given the variable number of aligned ESTs for each gene and their uneven coverage across its length, each splice site was classified in one or more tissues. For example, only three splices were classifiable in all 17 tissues considered and 991 in at least 10 tissues. We then investigated the classification concordance for introns labeled in five or more tissues. In most of cases (2993 introns, 62.2%) we detected an heterogeneous classification where the same intron was classified normal in some tissues and tumoral in others. For example, the first intron of PPP2R1A gene was exclusively expressed in the normal END tissue and in the tumoral BMA tissue (see Additional File 5). The remaining cases were homogeneously tumoral (1610, 33.5%) or normal (205, 4.3%).

We finally evaluated the classified introns on the basis of the spliceosomal machinery by which they are recognized (U2 or U12 spliceosome) without detecting any significant bias (data not shown).

The further step of our analysis was aimed to the identification of alternative N- and T-specific introns, based on their overlapping coordinates. Table 1 summarizes - for each considered tissue - the number of genes for which we were able to identify at least one tumor-associated splicing event not compatible (i.e. alternative) with a normal-associated one. The number of alternative tumoral and normal introns is also reported. We identified the highest number of genes [22] in the CNS, fitting the above mentioned criterion, related to the high number of available ESTs for this tissue.

To detect cancer (or normal) specific cassette exons, we selected those genes characterized by the simultaneous occurrence of at least one cancer specific splice site and at least one normal specific splice site defining overlapping introns (Table 1). The presence of a tumoral (or normal) cassette exon or mutually exclusive exons was defined by two cancer specific splice site pairs, alternative to one or two normal specific splice site pairs (or viceversa) (Figure 2).

We identified a total of 46 genes (for two of them, HLA-B and HLA-C, predictions were related to more than one anatomical group) with 88 condition-specific cassette or mutually exclusive exons, specifically 34 tumor-specific exons and 54 normal-specific exons which are spliced out in the tumoral isoforms. Interestingly, only one gene (SLC25A3) revealed mutually exclusive exons. For any of these cassette exons a differential expression in the normal and cancer conditions was previously reported.

We have investigated specific functional feature enrichments in these 46 genes by using DAVID [31] and Ingenuity Pathway Analysis [29]. Results are reported in Additional Files 6 and 7. Interestingly, the two top networks associated to this gene set concern cell function and maintenance and cell cycle. Also, an over-representation of cancer associated genes (23/46) can be detected. Concerning structural properties of the protein sequences encoded by cassette exons, a remarkable domain enrichment can be observed as 36/46 cassette exons overlapped annotated domains (78.3%) in contrast to the 54.4% (111,456/204,934) generally observed for protein coding internal exons in human genes.

Table 2 lists all genes for which we identified tumoral- or normal-specific cassette exons, including mutually exclusive exons (see Additional File 8 for a full list of the cassette exons).

Alternative splicing is widely recognized as a post-transcriptional regulatory mechanism which plays an important role in modulating gene expression in different tissues and developmental stages. Moreover, some alternatively spliced isoforms are associated with different diseases including cancer [5, 23, 24, 32]. Actually, mutations at the exon-intron junctions or affecting exonic or intronic splicing regulatory elements can contribute to the development of neoplastic states.

Histological information deriving from EST libraries annotation can be very useful in defining the expression profile of tumoral cells. The presence of a high number of tumoral ESTs and the absence of normal ESTs (or viceversa) mapping on a precise genomic region could mean that, in a cell, the presence/absence of a gene product may be cause or consequence of the tumoral pathology. The observation that the same gene may express many structurally and functionally different products, not only changed radically the traditional gene definition [33] but also extended the perspective of the analysis of differential expression in the normal and tumoral conditions beyond the gene level, taking into account alternative transcripts and proteins.

In this work, we have focused our attention on the study of alternative splicing aimed to the identification of cancer specific isoforms. Our method involves the characterization of splice sites with information deriving from EST libraries despite some intrinsic limitations of such data (34). Several studies of differential expression between the normal and tumoral conditions have been previously reported [23, 24, 35] through in silico analyses. Some predictions were also experimentally validated and collected in specialized databases, like ASAPII [13]. The method we developed for the classification of normal and tumoral introns differs from that used in ASAPII for two main reasons: 1) the statistics used; 2) the logic adopted for the definition of cancer specificity. Indeed, a serious limitation of the ASAPII database is that cancer specificity is detached from the tissue information as introns are classified cancer-specific regardless of the source tissue. In this way, as the Authors also report, some information can be lost or hidden by the general expression pattern of the gene, which may be tissue-specific. For example, in our analyses, for the KRT7 gene, the same splice site can be classified as normal- or cancer-specific depending on the tissue type considered (see Additional File 9). Such type of introns remain unclassified in the ASAPII database because it merges together ESTs from different tissues thus causing the flattening of the signal.

Another difference is represented by the fact that we define the splice site specificity in an absolute manner while in ASAPII database the tissue or histological specificity of each splice site is defined in comparison to another splice site. For the Tryptophanyl-tRNA synthetase gene (WARS), for example, the same splice site has been classified tumoral or normal depending on the other splice site with which it was compared to. In this way each splice site can not be univocally classified.

We approached the study of alternative splicing and its association to cancer, focusing, in the first part of the work, on the gene differential expression in normal and tumoral conditions among different tissues, in order to exclude from our candidate cancer specific splicing events, splice sites belonging to genes expressed in only one histological condition. We performed the classification analysis at intron level. The observation that CNS is the tissue with the highest number of tumor-specific splicing could imply some biological meaning or it could simply reflect the larger number of available normal and tumoral ESTs for the statistical validation.

Moreover, we specifically focused our attention on cassette exons to increase the reliability of statistical assessments, because of the necessity of a simultaneous occurrence of statistically validated tumor and normal-specific splicing sites, and to provide a collection of tumor biomarkers easily assayable and potentially suitable to be tested as therapeutic targets (e.g. using specific antibodies carrying cytotoxic agents).

Given the genome-wide approach (more than 18,000 genes analyzed), the number of significant results, a total of 88 cases, may not seem very high, but it must be recalled that we have used very stringent criteria which require simultaneous EST support for at least three splicing sites in both the normal and the tumoral status (see the data-flux flowchart in the Additional File 10). Furthermore, it should be considered that differential expression was based on EST counts from pooling normal and tumor libraries, thus loosing the replicate information. Forthcoming efforts using the huge amount of data provided by next-generation sequencing could suitably address this issue.

Our methodology has not identified all genes known to be characterized by the presence of tumor specific isoforms. For some genes like MDM2 and MDM4 [5], this was due to the insufficient number of available ESTs, either tumoral or normal, to define intron histological specificity. For CD44 [23] or TP53 [36] genes, we found cancer-specific splice sites in many tissues, but the number of normal ESTs was not sufficient to find significant normal-specific introns. Beside the small number of ESTs supporting introns prediction, another problem encountered has been the lack of ESTs annotation. The absence of information concerning tissue or histology associated to libraries led us to the exclusion of many EST sequences from the classification analysis. It is thus essential to reaffirm the importance of a good sequence annotation.

This methodology can be easily adapted to manage exon array hybridization data [4] and next-generation sequencing data [37], thus opening unprecedented opportunities for a thorough investigation of the expression pattern in human and other organisms.