Human pre-mRNA sequences are subjected to complex multi-stage modifications by splicing, where frequent variations in this process contribute to the proteome diversity. During splicing the intronic sequences are recognized and excised by the spliceosome, where the relatively short exonic sequences are joined together to form mature mRNA. The Splice Site (SS) signals at the intronic 5' end (donor) and 3' end (acceptor, polypyrimidine tract and the branch point) are necessary, but not sufficient for accurate and efficient exon recognition by the spliceosome [1, 2]. Additional exon-proximal elements are required for proper recognition of weakly defined or alternatively committed exons [3]. These cis-acting elements include a repertoire of Exonic Splicing Enhancers (ESEs) and Intronic Splicing Enhancers (ISEs) along with a number of Exonic Splicing Silencers (ESSs) and Intronic Splicing Silencers (ISSs). The evolutionary fine-tuned antagonism between enhancing and silencing elements leads to the proper splicing of human pre-mRNAs. Mutations disrupting cis-acting elements and SSs themselves, as well as mutations creating cryptic SSs and cis-acting factor binding sites can lead to severe diseases [4].

Mutations affecting alternative and constitutive splicing play a major role in human hereditary disorders [5]. More than 5,477 splicing mutations (as of July 2008) have been documented in the HGMD database [6], which makes this group of mutations one of the most frequent disease-causing alterations. Databases DBASS5 [7] and DBASS3 [8] contain 431 and 283 well annotated disease-causing aberrant splicing events, respectively. A clear understanding of elements affecting splicing could potentially aid diagnosis and development of novel therapeutic strategies [9, 10].

Since alternations in splicing are ubiquitous among human multi-exonic genes [11], it is important to understand the key regulators of this process. The 5'GC SSs, flanking <1% of human exons [12], were shown to play an important role in the genesis of alternative splicing in human genes [13] and were found to accumulate in mammalian lineage [14]. The majority of 5'GC SS sensors, i.e. computational procedures reporting how well an oligonucleotide would play a role of a SS, built up to date is based on weight matrices [12, 15]. Being an elusive signal, it is difficult to collect a representative learning set that would facilitate building a stronger model. The importance of proper modeling the 5'GC SS comes from the observation that some mutations documented in DBASS5 [7], such as IVS27+3_6dup(GGGT)(-96), IVS7+1G>T(-40), IVS9+1G>A(-45) and others trigger use of cryptic non-canonical 5'GC SS. Despite of the importance of this splicing signal only few splicing prediction methods, such as GeneSequer [16] and NetGene2 [17], are able to score non-canonical exons.

Human introns contain many decoy exons that are similar to authentic exons, but are never committed by the spliceosome and outnumber the real exons by an order of magnitude [18]. The mechanisms that allow accurate discrimination between decoy exons and their authentic counterparts are poorly understood. Codon sequence contained in coding exons have particular 3-periodic compositional biases [19] that allow gene finders, such as GenScan[20] and HMMgene[21], stitching putative coding exons in a frame-consistent fashion with high accuracy [22]. However, methods that rely on protein coding potential features experience severe performance loss when confronted with non-coding exons [23, 24]. On the other hand, human mutations frequently create de novo cryptic exons with no apparent coding potential leading to severe disorders caused by aberrant splicing [7, 8]. Therefore, tools are needed to explain the effects of mutations in terms of signals associated with splicing free of protein coding context [3].

Investigation whether prediction of SSs could be accomplished without relying on protein coding potential started with simple tools such as SpliceView[25] and GENIO[26]. The NetUTR[24] tool has been specifically constructed to predict SSs in 5' untranslated regions (UTRs), therefore addressing the problem of splicing prediction without relying on protein coding features. Maximum Entropy Sensor [27] has been found to be one of the most sensitive diagnostic methods predicting the effects of mutations in human genes [7, 8, 28]. ExonScan[29], a tool built around the exon definition model, combines the power of the Maximum Entropy Sensor with the Logarithm of Odds (LOD) biases associated with the previously reported ESEs [30], ESSs [29] and poly-G runs (known ISEs [31]). Recent CRYP-SKIP[32, 33] tool is based on multivariate logistic discrimination procedure that distinguishes the two aberrant splicing outcomes from DNA sequences. Bayesian SS sensor [23], shown to outperform the Maximum Entropy Sensor [34], is an integral part of the SpliceScan tool [23], built around the SS definition model supported by the enhancers predicted with the MHMMotif tool [23] and various other previously reported silencing and enhancing signals. The SpliceScan has been found to be especially efficient on the test set of short 5' UTR fragments.

We introduce a new tool SpliceScan II built around the exon definition model [1]. Unlike in previous SpliceScan[23] method, the new tool has option of displaying factors contributing to a score assigned to a specific exon isoform thus informing medical practitioners of possible changes in splicing commitment caused by polymorphic variants and mutations. We have used a much larger set of orthologous exons originating from 23 Tetrapoda organisms to train the new splicing model following an observation that the spliceosomal and cis-acting factors stay mostly intact across vertebrates [4, 35–38], where the genes encoding well-known RNA binding proteins involved in splicing regulation are enriched with ultraconserved elements [39]. The SpliceScan II tool is based on the Bayesian SS sensors, and uses the novel set of enhancer and silencer elements computationally predicted in Tetrapoda organisms [40]. Having a large collection of Tetrapoda orthologous exons we were able to collect learning set of 5'GC SSs, representative enough to train a new Bayesian 5'GC SS sensor, used in our tool. We compared the performance of our tool with other methods on gene fragments annotated in DBASS5 [7] and DBASS3 [8] and gene structures linked to Autism Spectrum Disorder (ASD). We further evaluated the method by predicting the effects on splicing for some of the previously reported polymorphisms associated with Alzheimer's and the Breast cancer, suggesting possible mechanism causing the disease predisposition associated with such variants.

Using the set of previously predicted cis-acting elements we were able to construct a splicing simulator capable of predicting exon score changes induced by mutations and polymorphic variants thus elucidating possible mechanism behind such variants leading to disorders caused by aberrant splicing.

Our tool performs favorably, compared to other splicing prediction methods, in context of genes linked to ASD. SpliceScan II provided more accurate prediction of aberrant splicing events, as documented in DBASS5 [7] and DBASS3 [8], compared to existing methods. Although the performance of our tool predicting the effect of mutations triggering an aberrant splicing is high compared to other methods, it could not be used as a general ab initio gene structural annotation method since the number of false positive predicted exons is high, as could be seen in Figure 2, though the fraction of reported false positives is comparable to what reported by other similar methods. Therefore, the most informative use of our method would evolve screening of polymorphic variants for possible splicing alternations in the context of known reference human gene structures. To accomplish this task we have created companion Autism Candidate Gene Map (ACGMAP) database http://www.meddean.luc.edu/node/375 that contains such structures and known alternative splicing variants for candidate ASD genes.

The reason the SpliceScan II is less specific (especially for higher sensitivity values) than previous SpliceScan tool [23] is in the nature of classification problem we address with new method. As could be seen in Figure 4, the SpliceScan uses simple probabilistic model of scoring putative 3' SS, where confidence of the putative 3' SS raises since two strong complement 5'SSs are located downstream. However, according to [7] 3'SS would unlikely to form exon with any of the 5' SSs located downstream in such a way since physiologically feasible 5'SS normally avoid strong competitors nearby. Indeed, SpliceScan scoring for both putative 5'SSs would be mediocre due to a conflict associated with their closeness. However, this observation does not help to predict which exonic isoform would be activated. To resolve this logical difficulty SpliceScan II makes all possible pairs of putative 3'SS and 5'SS located no further than 400 nt apart to predict possible SSs utilization. The number of pseudo exons formed this way outnumber the real exons by at least on order of magnitude [18], which turns in a harder classification task than simple SSs classification. For the weak SSs the number of putative exons to be classified is in excess of the number of weak SSs flanking them, which translates to a lower specificity compared to simpler SSs scoring methods.

Among the SNPs listed in dbSNP http://www.ncbi.nlm.nih.gov/projects/SNP/ as located within a gene locus, extended with 2 kb upstream and downstream flanks, only 1% are non synonymous variants [42]. The rarity of such SNPs suggested search for other causative variants affecting protein function through alternations in gene regulation, where disruption in splicing regulation seems a natural choice. Here we conducted search for causative alternations under realistic assumption that not all the SNPs associated with a disease are causative; chances are high that these SNPs were linked with causative variants located at the same haplotype. Moreover, associated variants may have different mechanism of compromising genome integrity such as creating missense/nonsense variants or affecting gene transcription regulatory elements. Nevertheless, we have been able to establish a number of potentially disease-causing splicing alternations.

Detected potentially detrimental exon score changes for disease associated SNPs are generally milder compared to the predicted changes associated with mutations annotated in DBASS databases [7, 8]. We did not predict any disease associated SNPs causing an annotated exon to disappear, an event that would most likely have highly detrimental consequences. Comparison to the predicted changes associated with the same sized control groups of randomly selected SNPs indicated that disease-causing SNPs have pronounced excess in the number of detected potentially splice-disrupting variants. Careful examination of factors contributing to an exon score variation could lead to a plausible explanation of causative mechanisms behind the disease associated SNPs.

The SpliceScan II is available online at http://splicescan2.lumc.edu/. The 5'GC SS sensor and the standalone SpliceScan II program could be found at http://www.wyomingbioinformatics.org/~achurban/.