GSNAP can align transcriptional reads that cross exon–exon junctions involving known or novel splice sites. For known splice sites, the program depends upon a user-provided set of splice sites, which belong to one of four categories: donors and acceptors on the plus genomic strand, and donors and acceptors on the minus genomic strand. Identification of novel splice sites is assisted by a probabilistic model, currently implemented as a maximum entropy model (Yeo and Burge, 2004 (link)), which uses frequencies of nucleotides neighboring a splice site to discriminate between true and false splice sites.
We use two methods for detecting splice junctions, one for short-distance and one for long-distance splicing. Short-distance splicing involves two splice sites that are on the same chromosomal strand, with the acceptor site being downstream of the donor site, within a user-specified parameter (default 200 000 nt). Short-distance splice junctions can be detected using a method similar to that for middle deletions, except that the distance allowed between candidate regions is much longer (Fig. 5B). As with middle indel detection, the positions of mismatches in the two regions determine whether a crossover area exists with the allowed number of mismatches (K − S), where S is the opening gap penalty for a splice. This crossover area is searched for donor and acceptor splice sites that are either known or supported by a splice site model at a sufficiently high probability. The probability score required is dependent on the length of short read sequence available for alignment in the exon region. When the aligned exon sequence is short, on the order of 12–20 nt, a relatively high probability score is needed. But when the aligned exon sequence is sufficiently long, more than 35 nt, only the expected dinucleotides at the intron end are needed.
For long-distance splicing, probability scores are also used to help find novel splice sites, although the required probability scores are higher for a given length of aligned sequence to compensate for the larger search space over the entire genome. To detect cases of long-distance splicing, the program identifies known or novel splice ends within single candidate regions, in the area delimited by the constraint level K of allowed mismatches (Fig. 5D). Candidate regions with donor and acceptor splice sites are then paired if they have the same breakpoint on the read, and have an acceptable number of total mismatches.
Reads that lie predominantly on one end of a splice junction may have too little sequence at the distant end to identify the other exon. Such alignments can still be reported by our program as partial splicing or ‘half intron’ alignments, if there is sufficient sequence on one end to determine a splice site, but insufficient sequence on the other end for the other site.