ISVASE: identification of sequence variant associated with splicing event using RNA-seq data

Background Exon recognition and splicing precisely and efficiently by spliceosome is the key to generate mature mRNAs. About one third or a half of disease-related mutations affect RNA splicing. Software PVAAS has been developed to identify variants associated with aberrant splicing by directly using RNA-seq data. However, it bases on the assumption that annotated splicing site is normal splicing, which is not true in fact. Results We develop the ISVASE, a tool for specifically identifying sequence variants associated with splicing events (SVASE) by using RNA-seq data. Comparing with PVAAS, our tool has several advantages, such as multi-pass stringent rule-dependent filters and statistical filters, only using split-reads, independent sequence variant identification in each part of splicing (junction), sequence variant detection for both of known and novel splicing event, additional exon-exon junction shift event detection if known splicing events provided, splicing signal evaluation, known DNA mutation and/or RNA editing data supported, higher precision and consistency, and short running time. Using a realistic RNA-seq dataset, we performed a case study to illustrate the functionality and effectiveness of our method. Moreover, the output of SVASEs can be used for downstream analysis such as splicing regulatory element study and sequence variant functional analysis. Conclusions ISVASE is useful for researchers interested in sequence variants (DNA mutation and/or RNA editing) associated with splicing events. The package is freely available at https://sourceforge.net/projects/isvase/. Electronic supplementary material The online version of this article (doi:10.1186/s12859-017-1732-7) contains supplementary material, which is available to authorized users.


Background
Alternative splicing is a normal phenomenon in eukaryotes and greatly increase the biodiversity of proteins. About 95% of multi-exonic genes are alternatively spliced in human [1]. The extreme example is the Drosophila Dscam gene, which produces thousands of protein isoforms by alternative splicing [2]. Classic pre-mRNA splicing is recognized and regulated by core splicing signals (5′ splice site (5′ ss), 3′ splice site (3′ ss), branch point sequence) and auxiliary sequences (splicing regulatory elements). Aberrant RNA splicing has become a common disease-causing mechanism, which can lead to hereditary disorders and cancers. Recent studies indicate that one third or a half of disease-causing mutations can affect RNA splicing [3,4]. Therefore, identification of sequence variant associated with splicing event (SVASE) becomes a meaningful procedure to illustrate the pathogenesis of diseases. Usually, sequence variant can result in aberrant splicing by disturbing regulatory element sequence or changing splice site [5]. For example, two sequence variants in splicing regulatory elements induce the aberrant splicing of BRCA2 exon 7 [6]. Moreover, RNA editing also can effect RNA splicing in transcriptome level [7].
Nowadays, RNA-seq has become a routine method for gene expression calling in multiple studies and can be also used to identify sequence variant and splicing event simultaneously [8,9]. However, there is only one bioinformatic tool (PVAAS) available for directly identifying genome-wide SVASE [10], which has some shortages, such as dependency on known splicing sites, only for novel splicing events, high false positive and long running time. Herein, we develop ISVASE, a suite of Perl scripts, to address the shortcomings of PVAAS and provide new functions for downstream analysis. The only necessary input files are genome sequence (FASTA format) and sequence alignment (BAM or SAM format) [11] files. The sequence alignment file must contain split-reads mapping result produced by software like GSNAP [12] and TopHat [13]. We also recommend users to provide known splicing events in GTF, GFF or BED format for junction shift event identification if concerned.

Implementation
The basic working principle of SVASE identification includes three main steps: (1) identify alternative splicing events; (2) identify sequence variants in specific splicing event using split-reads; and (3) evaluate the association between sequence variants and splicing events (see Fig. 1).
Based on sequence alignment result, ISVASE first filters mapped reads using stringent rule-dependent filters, such as low base quality (<Q30), low mapping quality (unpaired reads for paired-end data, PCR duplication, quality control, multiple mapping, mismatch, insertion and deletion) and short read length (<30 bp). Only splitreads with at least 8 bp anchor size in both parts of splicing event (junction) can be used to identify putative splicing event. Initially, splicing events with low read depth (<3) are discarded. Low abundant splicing events are also filtered out as background expression by applying Fisher's exact test to the putative splicing event and its related splicing events (sharing 5'ss or 3'ss). Here, ISVASE divides each splicing event into two independent parts based on 5'ss and 3'ss. ISVASE can remove known splicing events using annotation file in GTF, GFF or BED format by option "-k no". Although excellent software for sequence variant calling has existed such as GATK [14] and samtools [15], their results are hard to be used for SVASE calling, which needs to clarify specific sequence variants for unique splicing event. Thus, ISVASE adopts de novo sequence variant identification by only using junction-supporting split-reads. The observed sequence variant candidates are filtered by following criteria: reads depth (<3), alternative allele (ALT) supporting reads number (<3), ALT proportion (<0.1) and the significance of variant (p > 0.05, Fisher's exact test).
The practice of SVASE identification has a bit difference depending on whether the ALT frequencies are consistent between target splicing event and all related splicing events. We calculated the ALT frequencies for each sequence variants using reads of all splicing events and the target splicing event, respectively. If consistence, the association is assessed only using reads from target splicing event. Otherwise, total related reads are used. ISVASE applies same method as PVAAS to evaluate the significance of association. Besides, ISVASE assesses splicing signal by MaxEntScan [16] and identifies junction shift events to reduce the false positive of splicing event calling. Furthermore, DNA mutation and/or RNA editing profiles (like dbSNP [17], DARNED [18], RADAR [19] or user provided DNA mutation or RNA editing Fig. 1 Schematic diagram of the ISVASE software. a Identify splicing variants in RNA-seq data. All splicing variants can be divided into four types according to relationship between target splicing variant (red colour) and other splicing variants (from left to right): (i) unique splicing variant; (ii) splicing variants with same junction start; (iii) splicing variants with same junction end; and (iv) splicing variants with same junction start or end. b Identify sequence variants for each splicing variant and all related splicing variants. To handle all splicing variant types, we identify sequence variants for two parts of splicing separately. In the left part, for junctions with orange, yellow and red colour, the all related splicing variants should be three (all these junctions); however, for junctions with green and blue colour, the total junction is one (itself). Similarly, in the right part, junctions with red, green and blue colour have three all related splicing variants while junctions with orange and yellow colour only has one related junction (itself). c Identify associations. This step includes three significant judgements for sequence variants, junction existence and association between sequence variants and junctions, respectively. The example shown two junctions with same junction end. For junction one (top), two sequence variants are identified (left G(ref)-> C(alt) and right G(ref)-> A(alt)). In sequence variant significant judgement, left is filtered (p value = 1) while right passes the test (p value = 0.0476). In junction significant judgement and association judgement, p value of top junction is 0.0128 (significant) and 0.0070 (significant) respectively. Dashed lines represent gaps in the alignment sites) can be used to assign the source type of sequence variants. ISVASE outputs the detailed statistical results with figures and tables. ISVASE also extracts the flanking sequence for sequence variants, which can be used to predict exonic splicing enhancer (ESE) motifs using tools like ESEfinder [20] and Human Splicing Finder [21]. The output of identified SVASEs can be accepted by ANNOVAR [22] and SnpEff [23] for further functional analysis like KEGG pathway and Gene Ontology. The code of ISVASE was written using Perl (v5.18.4), the figures were created by R (v3.1.2) while the sequence alignment file was operated by samtools (v1.2).

Results and Discussion
To demonstrate the functionality of ISVASE and compare with PVAAS, PVAAS testing data (downloaded from website http://pvaas.sourceforge.net/) was used. PVAAS (v0.1.5) identified 8 SVASEs (belonging to new splicing events), while ISVASE obtained 172 SVASEs and 14 of them were new splicing events (Table 1, Additional files 1 and 2). Two software only share one SVASE, which probably is genuine according to dbscSNV [24]. Among other 7 PVAAS unique SVASEs, 1 SVASE has a low ALT ratio (<=0.01), 1 SVASE is supported by un-split reads and remaining 5 SVASEs are identified only by a small part of   (Table 3).
To further reveal the advantage of ISVASE, we also test another real data set with 4 RNA-seq samples for human glioblastoma cell line U87MG (SRR388226 and SRR388227 are control samples and SRR388228 and SRR388229 are ADAR knockdown samples) [25]. The raw data was trimmed by Trimmomatic [26] and aligned by GSNAP (only concordant mapping results were used for downstream analysis) [12]. Using ISVASE, 134 and 120 SVASEs (87 common) were obtained for control data, while 187 and 168 SVASEs (119 common) for knockdown data in new splicing events. If considering all splicing events, 2105 and 2298 common SVASEs were identified in control and knockdown data ( Table 1, Additional files 3, 4, 5, and 6). In each sample, at most three SVASEs belonging to RNA editing sites in RADAR database were detected (totally four SVASEs belonging to RADAR database), and more than 82% SVASEs existed in dbSNP. In comparison, PVAAS got 61 and 63 SVASEs (28 common) for control data, while 93 and 89 SVASEs (31 common) for knockdown data ( Table 1, Additional files 7, 8, 9, and 10). In PVAAS result, there wasn't any SVASE belonging to RNA editing sites in RADAR database and at most 27% SVASEs existed in dbSNP. These results indicated that PVAAS has higher false positive rate comparing with ISVASE ( Table 2). Using repeat samples, we also found that PVAAS has lower consistency rate comparing with ISVASE (about 47% vs. about 83%) ( Table 2). Moreover, for each sample, ISVASE showed an advantage of running time to PVAAS (about 3 h vs. 14.34 h) ( Table 3).
The SVASEs identified by ISVASE can be used for downstream analysis easily. For example, we used 65 common SVASEs in new splicing events from the above four samples to do further analysis. We annotated these SVASEs by ANNOVAR and found 28 related genes (Additional file 11). Among them, 20, 9 and 8 SVASEs located in HLA, HCG4B and AHNAK2 genes. HLA genes play important roles in tumor immune surveillance and escape, and HCG4B gene is a pseudogene of HLA complex group. AHNAK2 gene is associated with  calcium channel proteins and its exon 7 size is almost 18 kb. We found 8 SVASEs associated with 5 new splicing events inside the exon 7. Gene Ontology enrichment analysis found these 28 genes are significantly enriched in cancer related functions, such as antigen processing and presentation, response to type I interferon and interferon-gamma (Table 4). We also used ESEfinder to detect ESE motifs and found 57 of 65 SVASEs located in predicted ESE motifs. This result indicates most of SVASEs perform their function possibly by influencing ESE motifs of splicing events. Moreover, SVASEs have some basic characteristics (using SRR388226 data as an example), such as high proportion of canonical splicing signal GT-AG (or reverse complement CT-AC), similar signal scores for splice Fig. 2 The characteristics of SVASEs between novel and all SVASE sites in sample SRR388226. The density of junction reads number, the bar plot of junction number for different junction splicing signals, the boxplot of junction reads number distribution for different junction splicing signals, the density of splicing signal score for variant replaced sequence and reference sequence, the histogram plot of distances between sequence variant and exon 5′ side, the histogram plot of distances between sequence variant and exon 3′ side, the boxplot of distance distribution between sequence variant type and junction breakpoint, and the bar plot of sequence variant number for different sequence variant types are shown for SVASEs located in new splicing events (the upper half) and all splicing events (the lower half) sites with reference or alternative allele, tendency to junction breakpoints, and high frequency of A-> G/T-> C and G-> A/C-> T transitions (58.96% in new splicing events and 75.13% in all splicing events) (Fig. 2).