Logo of narLink to Publisher's site
Nucleic Acids Res. Oct 2010; 38(18): e178.
Published online Aug 27, 2010. doi:  10.1093/nar/gkq622
PMCID: PMC2952873

MapSplice: Accurate mapping of RNA-seq reads for splice junction discovery

Abstract

The accurate mapping of reads that span splice junctions is a critical component of all analytic techniques that work with RNA-seq data. We introduce a second generation splice detection algorithm, MapSplice, whose focus is high sensitivity and specificity in the detection of splices as well as CPU and memory efficiency. MapSplice can be applied to both short (<75 bp) and long reads (≥75 bp). MapSplice is not dependent on splice site features or intron length, consequently it can detect novel canonical as well as non-canonical splices. MapSplice leverages the quality and diversity of read alignments of a given splice to increase accuracy. We demonstrate that MapSplice achieves higher sensitivity and specificity than TopHat and SpliceMap on a set of simulated RNA-seq data. Experimental studies also support the accuracy of the algorithm. Splice junctions derived from eight breast cancer RNA-seq datasets recapitulated the extensiveness of alternative splicing on a global level as well as the differences between molecular subtypes of breast cancer. These combined results indicate that MapSplice is a highly accurate algorithm for the alignment of RNA-seq reads to splice junctions. Software download URL: http://www.netlab.uky.edu/p/bioinfo/MapSplice.

INTRODUCTION

Alternative splicing is a fundamental mechanism that generates transcript diversity. Particular combinations of cis-acting sequences, trans-acting splicing regulators and histone modifications contribute to differential exon usage among diverse cell types (1,2). Through shuffling of exons, splice sites and untranslated regions can drastically alter the cellular function of proteins (3,4). Notably, SNPs have been linked to changes in transcript isoform proportions among different individuals (5). In some cases, rare mutations that alter splicing patterns have been linked to disease (6–9). Thus, transcriptome profiling should comprise a comprehensive survey of alternative splicing.

Microarrays were the first technology to enable global assessment of alternative splicing (10–13). Oligonucleotides designed to span two adjacent exons can be used to measure the abundance of splice junctions. However, these splice junction probes only interrogate a predefined set of transcript isoforms. Due to the large number of hypothetical exon–exon combinations, microarrays are not efficient at discovering novel transcript isoforms.

Deep transcriptome sequencing provides sufficient read counts to measure relative proportions of transcript isoforms, as well as to discover new isoforms (1,14–17). Several high-throughput technologies currently sample short sequence tags, typically <200 bp in size. The accurate mapping of sequence tags that span splice junctions is the foundation for transcript isoform reconstruction (18,19). One approach relies on existing transcript annotations to create a database of potential splice junction sequences. Similar to the above limitation with microarrays, the construction of predefined alignment databases limits the set of possible splice junctions interrogated.

Recently methods have been developed to find novel splice junctions from short sequence tags. The pioneering QPALMA algorithm adopted a machine learning algorithm to predict splice junctions from a training set of positive controls (20). The TopHat algorithm constructs candidate splice junctions by pairing candidate exons and evaluating the alignment of reads to such candidates (21). SpliceMap is another method that uses splice site flanking bases in locating potential splice sites (22).

We introduce the MapSplice algorithm to detect splice junctions without any dependence on splice site features. This enables MapSplice to discover non-canonical junctions and other novel splicing events, in additional to the more common canonical junctions. MapSplice can be generally applied to both short and long RNA-seq reads. In addition, MapSplice leverages the quality and diversity of read alignments that include a given splice to increase specificity in junction discovery. As a result, MapSplice demonstrates high specificity and sensitivity. Performance results are established using synthetic data sets and validated experimentally.

We have used MapSplice to investigate significant differences in alternative splicing between a set of basal and luminal breast cancer tissues. Experimental validation of 20 exon skipping events by quantitative RT–PCR (qRT–PCR) correctly identified isoform proportions that are highly correlated (Pearson's correlation = 0.86) with their estimates based on splice junctions. Splice junctions also recapitulated the difference between molecular subtypes of breast cancer. On a global level, the proportion of splice junctions in various categories of alternative splicing was concordant with a previous RNA-seq study.

MATERIALS AND METHODS

The goal of MapSplice is to find the exon splice junctions present in the sampled mRNA transcriptome, and to determine the most likely alignment of each mRNA sequence tag to a reference genome. Each tag corresponds to a number of consecutive nucleotides read from an mRNA transcript, where the length of the tag is determined by the protocol and the sequencing technology. For example, the Illumina Genome Analyzer IIx generates over 20M tags of size up to 100 bp per sequencing lane.

MapSplice operates in two phases to achieve its goal. In the ‘tag alignment' phase, candidate alignments of the mRNA tags to the reference genome G are determined. Tags with a contiguous alignment fall within an exon and can be mapped directly to An external file that holds a picture, illustration, etc.
Object name is gkq622i2.jpg but tags that include one or more splice junctions require a gapped alignment, with each gap corresponding to an intron spliced out during transcription. Since multiple possible alignments may be found, the result of this phase is, in general, a set of candidate alignments for each tag.

In the ‘splice inference phase,’ splice junctions that appear in the alignments of one or more tags are analyzed to determine a splice significance score based on the quality and diversity of alignments that include the splice. The purpose of this phase is to reject spurious splices and to provide a basis for choosing the most likely alignment for each tag based on a combination of alignment quality and splice significance. An overview of the algorithm can be found in Figure 1. The two phases are described in the following two sections.

Figure 1.
An overview of the MapSplice pipeline. The algorithm contains two phases: tag alignment (Step 1–Step 4) and splice inference (Step 5–Step 6). In the ‘tag alignment' phase, candidate alignments of the mRNA tags to the reference ...

Tag alignment

Let Θ be the set of tags and let m be the tag length. A tag An external file that holds a picture, illustration, etc.
Object name is gkq622i4.jpg has an ‘exonic alignment' if it can be aligned in its entirety to a consecutive sequence of nucleotides in G. T has a ‘spliced alignment' if its alignment to G requires one or more gaps.

MapSplice identifies candidate tag alignments in three steps. First, tags are partitioned into consecutive short segments and an exonic alignment to G is attempted for each segment. In the second step, segments that do not have an exonic alignment are considered for spliced alignment using a splice junction search technique that starts from neighboring segments already aligned. In the final step, segment alignments for a tag are merged to find candidate overall alignments for each tag. The details of the steps follow next.

Step 1: partition tags into segments

Tags in Θ of length m are partitioned into n consecutive segments of length k where An external file that holds a picture, illustration, etc.
Object name is gkq622i12.jpg. Typically k is 20–25 for tags of length 50 or greater. As k is decreased, the chance that a segment contains one or more splice junctions decreases correspondingly, but the chance for multiple spurious alignments of the segment increases. The segments making up a tag T are labeled An external file that holds a picture, illustration, etc.
Object name is gkq622i15.jpg where the number of segments An external file that holds a picture, illustration, etc.
Object name is gkq622i16.jpg

Step 2: exonic alignment of segments

Exonic alignment of segments can be performed using fast approximate aligners such as Bowtie (23), and BWA (24), or aligners with more general error tolerance models such as SOAP2 (25), BFAST (26) and MAQ (27). For each segment An external file that holds a picture, illustration, etc.
Object name is gkq622i17.jpg of tag T, let An external file that holds a picture, illustration, etc.
Object name is gkq622i18.jpg be the number of possible exonic alignments of An external file that holds a picture, illustration, etc.
Object name is gkq622i19.jpg to the genome, determined using one of the algorithms mentioned above with an error tolerance of An external file that holds a picture, illustration, etc.
Object name is gkq622i20.jpg mismatches. When An external file that holds a picture, illustration, etc.
Object name is gkq622i21.jpg the segment has a unique exonic alignment. When An external file that holds a picture, illustration, etc.
Object name is gkq622i22.jpg the segment has multiple alignments, each of which is considered in the subsequent steps.

Step 3: spliced alignment of segments

If An external file that holds a picture, illustration, etc.
Object name is gkq622i23.jpg, segment An external file that holds a picture, illustration, etc.
Object name is gkq622i24.jpg does not have an exonic alignment. One possible reason is that it may have a gapped alignment crossing a splice junction. In general, if the minimum exon length is at least 2k, then for every pair of consecutive segments in T at least one segment should have an exonic alignment. Therefore, the alignment of segment An external file that holds a picture, illustration, etc.
Object name is gkq622i25.jpg is localized to the alignments of its neighbors. The following two techniques are used to find a spliced alignment for segments, and are illustrated in Figure 2.

Figure 2.
A portion of an mRNA transcript sampled by tag An external file that holds a picture, illustration, etc.
Object name is gkq622i26.jpg consists of the 3′ end of exon 1, all of exon 2 and the 5′ end of exon 3. An external file that holds a picture, illustration, etc.
Object name is gkq622i27.jpg is split into segments t1,…, tn each of length An external file that holds a picture, illustration, etc.
Object name is gkq622i28.jpg to identify the alignment of An external file that holds a picture, illustration, etc.
Object name is gkq622i29.jpg to the genome. Provided ...

If An external file that holds a picture, illustration, etc.
Object name is gkq622i37.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i38.jpg both have exonic alignments, then we perform a ‘double-anchored' spliced alignment of An external file that holds a picture, illustration, etc.
Object name is gkq622i39.jpg for all combinations of exonic alignments of An external file that holds a picture, illustration, etc.
Object name is gkq622i40.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i41.jpg. If only one neighboring segment An external file that holds a picture, illustration, etc.
Object name is gkq622i42.jpg has an exonic alignment, then we perform a ‘single-anchored' alignment starting from the An external file that holds a picture, illustration, etc.
Object name is gkq622i43.jpg possible alignments of An external file that holds a picture, illustration, etc.
Object name is gkq622i44.jpg.

  1. Double-anchored spliced alignment: the spliced alignment of An external file that holds a picture, illustration, etc.
Object name is gkq622i45.jpg to the genomic interval between anchors An external file that holds a picture, illustration, etc.
Object name is gkq622i46.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i47.jpg need only consider the An external file that holds a picture, illustration, etc.
Object name is gkq622i48.jpg possible positions of the splice junction x within An external file that holds a picture, illustration, etc.
Object name is gkq622i49.jpg and minimize alignment mismatch.
    Formally, the ‘Hamming distance' An external file that holds a picture, illustration, etc.
Object name is gkq622i50.jpg between two equal length sequences S and T is defined as the number of corresponding positions with mismatching bases. We define the spliced alignment between the segment An external file that holds a picture, illustration, etc.
Object name is gkq622i52.jpg and the genomic interval An external file that holds a picture, illustration, etc.
Object name is gkq622i53.jpg as
    equation image
    which yields the optimal position x of the splice junction in t that gives the best spliced alignment to the given genomic interval. The splice junction x defines the intron as An external file that holds a picture, illustration, etc.
Object name is gkq622i55.jpg
    To find the spliced alignment for An external file that holds a picture, illustration, etc.
Object name is gkq622i56.jpg between two aligned segments, let An external file that holds a picture, illustration, etc.
Object name is gkq622i57.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i58.jpg be the leftmost genomic coordinates in the alignment of An external file that holds a picture, illustration, etc.
Object name is gkq622i59.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i60.jpg respectively, and compute spliced–alignAn external file that holds a picture, illustration, etc.
Object name is gkq622i61.jpg
    In case the alignment cost for the splice junction exceeds the error tolerance threshold An external file that holds a picture, illustration, etc.
Object name is gkq622i62.jpg the alignment for An external file that holds a picture, illustration, etc.
Object name is gkq622i63.jpg fails. If there exists more than one splice junction position with minimum cost for An external file that holds a picture, illustration, etc.
Object name is gkq622i64.jpg multiple alignments are recorded for An external file that holds a picture, illustration, etc.
Object name is gkq622i65.jpg
  2. Single-anchored spliced alignment: in the case of a single anchor An external file that holds a picture, illustration, etc.
Object name is gkq622i66.jpg upstream of the unaligned An external file that holds a picture, illustration, etc.
Object name is gkq622i67.jpg we conduct a search for An external file that holds a picture, illustration, etc.
Object name is gkq622i68.jpg the h-base suffix of An external file that holds a picture, illustration, etc.
Object name is gkq622i69.jpg in the genomic region downstream from An external file that holds a picture, illustration, etc.
Object name is gkq622i70.jpg Similarly, in the case of a single anchor An external file that holds a picture, illustration, etc.
Object name is gkq622i71.jpg downstream from An external file that holds a picture, illustration, etc.
Object name is gkq622i72.jpg the search is for the h-base prefix An external file that holds a picture, illustration, etc.
Object name is gkq622i74.jpg of An external file that holds a picture, illustration, etc.
Object name is gkq622i75.jpg in the region upstream from An external file that holds a picture, illustration, etc.
Object name is gkq622i76.jpg In either case this search is limited in range by a parameter D, the maximum intron size for single anchor search, typically set to 50 000 bp.
    All single anchored alignments can be resolved with a single traversal of (the expressed portion of) the genome using a sliding window of size D. An h-mer index is maintained during this traversal, mapping occurrences of an h-mer An external file that holds a picture, illustration, etc.
Object name is gkq622i77.jpg to the downstream anchor An external file that holds a picture, illustration, etc.
Object name is gkq622i78.jpg within distance D and occurrences of an h-mer An external file that holds a picture, illustration, etc.
Object name is gkq622i79.jpg to the upstream anchor An external file that holds a picture, illustration, etc.
Object name is gkq622i80.jpg within distance D. As the window moves, new entries are added as anchors come within range, and old entries are deleted as anchors fall out of range.
    When the h-mer at the current coordinate c in the genome scan is mapped to a downstream segment An external file that holds a picture, illustration, etc.
Object name is gkq622i81.jpg spliced-alignAn external file that holds a picture, illustration, etc.
Object name is gkq622i82.jpg gives the best spliced alignment which is recorded if it is within the segment error threshold An external file that holds a picture, illustration, etc.
Object name is gkq622i83.jpg. Similarly, when the h-mer is mapped to an upstream segment An external file that holds a picture, illustration, etc.
Object name is gkq622i85.jpg, we record spliced-alignAn external file that holds a picture, illustration, etc.
Object name is gkq622i86.jpg if it is within the segment error threshold An external file that holds a picture, illustration, etc.
Object name is gkq622i87.jpg.
  3. Spliced alignment in the presence of small exons: if an exon shorter than An external file that holds a picture, illustration, etc.
Object name is gkq622i88.jpg is included in a transcript, it is possible that two adjacent segments An external file that holds a picture, illustration, etc.
Object name is gkq622i89.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i90.jpg both include a splice junction so that neither can be aligned continuously within an exonic region. If the exon is shorter than k, even a single segment might include more than one splice junction. The following approach allows us to detect exons with size less than An external file that holds a picture, illustration, etc.
Object name is gkq622i92.jpg.
    Assume S is a sequence of one or two missed segments between two anchors An external file that holds a picture, illustration, etc.
Object name is gkq622i94.jpg that cannot be successfully aligned in the previous steps, potentially due to short exons. We divide S into a sequential set of h-mers and index S with these h-mers. By extending the sequential scan of the genome used in single-anchored spliced alignment, h-mers on the reference genome in An external file that holds a picture, illustration, etc.
Object name is gkq622i100.jpg can all be searched simultaneously. When a match exists, two double-anchored spliced alignments will be performed: one is between a and the 5′-site of the h-mer alignment; and the other is between the 3′-site of h-mer alignment and b.
    According to the pigeon-hole principle, if the exon is no shorter than An external file that holds a picture, illustration, etc.
Object name is gkq622i105.jpg, one of the h-mers in the unaligned segments will fall within an exon and thus trigger the subsequent spliced alignments. Therefore, this method is guaranteed to detect small exons longer than An external file that holds a picture, illustration, etc.
Object name is gkq622i107.jpg and possibly detect shorter exons. The typical h-mer size is 6–8 bps. When exons are shorter than An external file that holds a picture, illustration, etc.
Object name is gkq622i108.jpg, the chances of finding a spliced alignment decrease. Reducing h will lead to an increasing number of spurious matches that will be difficult to filter out.

Step 4: merging segment alignments

The assembly of a complete tag alignment from individual alignments of its segments is straightforward if each segment is aligned uniquely and connects to its neighboring segments without a gap. However, a given segment An external file that holds a picture, illustration, etc.
Object name is gkq622i110.jpg may be aligned at multiple locations. In this case, the possible combinations of alignments must be searched to find the best overall alignment for the tag.

Let An external file that holds a picture, illustration, etc.
Object name is gkq622i111.jpg be the set of alignments for segment An external file that holds a picture, illustration, etc.
Object name is gkq622i112.jpg and when An external file that holds a picture, illustration, etc.
Object name is gkq622i113.jpg, let An external file that holds a picture, illustration, etc.
Object name is gkq622i114.jpg be the An external file that holds a picture, illustration, etc.
Object name is gkq622i115.jpg alignment of An external file that holds a picture, illustration, etc.
Object name is gkq622i116.jpg, where An external file that holds a picture, illustration, etc.
Object name is gkq622i117.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i118.jpg. In principle there exist An external file that holds a picture, illustration, etc.
Object name is gkq622i119.jpg different combinations of alignments, but most can be ruled out by a simple coherence test based on contiguity of consecutive segment alignments.

Two adjacent segments An external file that holds a picture, illustration, etc.
Object name is gkq622i120.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i121.jpg with exonic alignment that are not contiguous on the genome are checked for a splice junction between the two segments using the double-anchored spliced alignment method. This procedure also corrects inaccurate splice points due to the error tolerance in the alignment of An external file that holds a picture, illustration, etc.
Object name is gkq622i122.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i123.jpg.

For each assembly of segments that yields a candidate alignment for T, we compute its ‘mismatch score', a modified Hamming distance between T and its alignment to the genome An external file that holds a picture, illustration, etc.
Object name is gkq622i125.jpg.

The mismatch score takes into account base call qualities when available. A poor quality base call can improve the score when associated with a mismatched base, but can also decrease the score when associated with a matched base (28). The base call quality for a given base x in the overall alignment of a tag T can be converted to a probability p that x was called incorrectly, and so the expected mismatch An external file that holds a picture, illustration, etc.
Object name is gkq622i130.jpg in the alignment of base x to base y in the genome is given by

equation image
(1)

where An external file that holds a picture, illustration, etc.
Object name is gkq622i133.jpg is the probability of base x in the background distribution of nucleotides. We assume a uniform distribution for nucleotides hence An external file that holds a picture, illustration, etc.
Object name is gkq622i135.jpg for all x. Thus, given a proposed alignment of An external file that holds a picture, illustration, etc.
Object name is gkq622i137.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i138.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i139.jpg the probability that nucleotide An external file that holds a picture, illustration, etc.
Object name is gkq622i140.jpg was called incorrectly, the expected mismatch is An external file that holds a picture, illustration, etc.
Object name is gkq622i141.jpg

The candidate alignment is retained if An external file that holds a picture, illustration, etc.
Object name is gkq622i142.jpg otherwise it is discarded. Note that while each segment was aligned allowing An external file that holds a picture, illustration, etc.
Object name is gkq622i143.jpg mismatches, the overall tag alignment only allows a total of An external file that holds a picture, illustration, etc.
Object name is gkq622i144.jpg expected mismatches. We define the quality of the alignment to be An external file that holds a picture, illustration, etc.
Object name is gkq622i145.jpg

Splice junction inference

Splice junction alignments introduce a multiplicity of ways in which a tag may be split into pieces, each of which may be separately aligned to the genome. For a given tag, at most one of these is the true alignment. Splice inference leverages the extensive sampling of splice junctions by tags to compute a junction quality that can be used to distinguish true splice junctions from spurious splice junctions and to determine the best alignment among the remaining candidate alignments of a tag.

Step 5: splice junction quality

For a given splice An external file that holds a picture, illustration, etc.
Object name is gkq622i146.jpg where An external file that holds a picture, illustration, etc.
Object name is gkq622i147.jpg is the last coordinate of the donor exon and An external file that holds a picture, illustration, etc.
Object name is gkq622i148.jpg is the first coordinate of the acceptor exon, we consider the set An external file that holds a picture, illustration, etc.
Object name is gkq622i149.jpg of tags that include a splice junction for J in a candidate alignment. We define two statistical measures on An external file that holds a picture, illustration, etc.
Object name is gkq622i151.jpg: the ‘anchor significance' An external file that holds a picture, illustration, etc.
Object name is gkq622i152.jpg determined by the alignment in An external file that holds a picture, illustration, etc.
Object name is gkq622i153.jpg that maximizes significance as a result of long anchors on each side of the splice junction, and the ‘entropy' An external file that holds a picture, illustration, etc.
Object name is gkq622i154.jpg measured by the diversity of splice junction positions in An external file that holds a picture, illustration, etc.
Object name is gkq622i155.jpg

  1. Anchor significance of a splice junction: a tag that includes a splice junction has some number of contiguous bases aligned on either side of the splice site. An alignment with a short anchor on one side has low confidence, since we expect it to be easy to find other occurrences of the sequence of nucleotides found in the short anchor each of which might equally well be the correct target. We define the anchor significance of a splice J in a tag An external file that holds a picture, illustration, etc.
Object name is gkq622i157.jpg as follows. Let An external file that holds a picture, illustration, etc.
Object name is gkq622i158.jpg be the maximal contiguous sequence of bases in T with exonic alignment ending at coordinate An external file that holds a picture, illustration, etc.
Object name is gkq622i160.jpg in the genome, and let An external file that holds a picture, illustration, etc.
Object name is gkq622i161.jpg be the maximal contiguous sequence of bases in T with exonic alignment starting from coordinate An external file that holds a picture, illustration, etc.
Object name is gkq622i163.jpg These are the two anchors, each of which has at least one alignment (the one in T). The expected number of alignments in the genome for an anchor An external file that holds a picture, illustration, etc.
Object name is gkq622i164.jpg is therefore given by
    equation image
    Here, we model the genome as a sequence of independent random variables with uniform distribution over A, C, T, G, so that the chance that a length n sequence aligns at a given coordinate is simply 4−n. For double-anchor alignments, the search space is effectively the entire genome of length N. For single anchor alignments, we only consider occurrences within distance D.
    Since we assume that only one of the potential alignments is correct and the rest are spurious, the chance of a spurious alignment of anchor An external file that holds a picture, illustration, etc.
Object name is gkq622i166.jpg is An external file that holds a picture, illustration, etc.
Object name is gkq622i167.jpg Thus the log-transformed significance of anchor An external file that holds a picture, illustration, etc.
Object name is gkq622i168.jpg is An external file that holds a picture, illustration, etc.
Object name is gkq622i169.jpg The anchor alignment of junction J in T is only as significant as the anchor with least confidence, hence is An external file that holds a picture, illustration, etc.
Object name is gkq622i172.jpg
    The anchor significance of junction J over all occurrences in An external file that holds a picture, illustration, etc.
Object name is gkq622i174.jpg is the occurrence with greatest anchor significance:
    equation image
  2. Entropy: in principle, the RNA-seq protocol samples each transcript uniformly, so that the position of a true splice junction J within An external file that holds a picture, illustration, etc.
Object name is gkq622i176.jpg is expected to be uniformly distributed on An external file that holds a picture, illustration, etc.
Object name is gkq622i177.jpg provided the sampling is sufficiently deep and the splice junction is not too close to the end of the transcript. To measure the uniformity of the sampling, we apply Shannon maximum entropy to the distribution of splice junction positions in An external file that holds a picture, illustration, etc.
Object name is gkq622i178.jpg for splice J. Let An external file that holds a picture, illustration, etc.
Object name is gkq622i180.jpg with An external file that holds a picture, illustration, etc.
Object name is gkq622i181.jpg be the frequency of occurrence of a splice junction J at position i within An external file that holds a picture, illustration, etc.
Object name is gkq622i184.jpg The Shannon entropy can be measured as
    equation image
    The higher the Shannon entropy, the closer the distribution is to uniform, and therefore the higher the chance that the junction is part of some transcript that was sampled uniformly.
  3. Combined Metric: the combined metric An external file that holds a picture, illustration, etc.
Object name is gkq622i185.jpg is the posterior probability that junction J is a true junction determined using Bayesian regression. The observed data of An external file that holds a picture, illustration, etc.
Object name is gkq622i187.jpg are the entropy and anchor significance of J within An external file that holds a picture, illustration, etc.
Object name is gkq622i189.jpg and the average quality of read alignments that include J.
    equation image
    We apply linear regression to obtain the best configuration of An external file that holds a picture, illustration, etc.
Object name is gkq622i191.jpg An external file that holds a picture, illustration, etc.
Object name is gkq622i192.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i193.jpg that achieves the maximum sensitivity and specificity in junction classification.

Step 6: best alignment for tags

For each tag T, we select the candidate alignment An external file that holds a picture, illustration, etc.
Object name is gkq622i194.jpg that achieves the highest score when combining alignment quality from Step 4 and junction quality from Step 5.

Synthetic data generation for validation

To evaluate the sensitivity and specificity of MapSplice, we generated synthetic data sets of tags derived from transcripts cataloged in the Alternative Splicing and Transcript Diversity (ASTD) database (29).

This database collects full-length transcripts illustrating alternative splicing events in genes from human, mouse and rat. A synthetic `transcriptome' is generated by randomly selecting genes and expression levels according to an empirical distribution of tags per gene observed in Ref. (1). Within a gene, transcripts are selected at random following various submodels that determine expression level of individual transcripts relative to the overall gene. The synthetic transcriptome characterized in this fashion is then sampled to yield two synthetic RNA-seq data sets. The noise-free data set samples the transcripts exactly and the resultant tags align to the reference genome exactly to model single-nucleotide variations in the data base transcripts. The noisy data set introduces mutations into base calls following empirical Illumina base call quality profiles. The resulting data sets mimic the observed distribution of errors in tags in Ref. (30).

Experimental validation by qRT–PCR

Total RNA isolated from MCF-7 and SUM-102 cells was reverse transcribed using the High-Capacity cDNA Reverse Transcription Kit with RNase inhibitor (Applied Biosystems, Foster City, CA, USA) as per manufacturer’s instructions. Relative expression levels of the transcripts of interest were determined by qRT–PCR on the Applied Biosystems 7300 Real Time PCR System with premade or custom TaqMan Gene Expression Assays (Applied Biosystems, Foster City, CA, USA) containing primers flanking the splice sites of interest and FAM/MGB-labeled oligonucleotide probes. PCR reactions were carried out as per manufacturer’s instructions. cDNA equivalent to 100 ng of total RNA was amplified with 1 μl of TaqMan assay in Gene Expression master mix in a total volume of 20 μl. Each assay was performed in triplicate. Thermal Cycling conditions were as follows: 50°C for 2 min, 95°C for 10 min, 40 cycles of 95°C for 15 s and 60°C for 1 min. Ct values were determined in the manufacturer’s software, data was further analyzed in Excel utilizing comparative Ct method. For comparisons of relative expression levels between the two cell lines, Ct values for the transcripts of interest were first normalized to those of HPRT1.

RESULTS

Junction inference

We constructed a synthetic noise-free RNA-seq data set with 20M 100 bp tags sampling 46 311 distinct transcripts from the ASTD. The tags were aligned to the reference genome (hg18) using the MapSplice algorithm steps 1–4 with An external file that holds a picture, illustration, etc.
Object name is gkq622i195.jpg An external file that holds a picture, illustration, etc.
Object name is gkq622i196.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i197.jpg To establish the training data set containing both true and false junctions, no restrictions on splice site flanking sequences or maximum intron size were enforced.

We randomly selected 10K true junctions and 10K false junctions as a training set and analyzed the three different junction classification metrics utilized in Step 5 of MapSplice: alignment quality entropy and anchor significance, as well the combined metric obtained by linear regression of the first three metrics. Five-fold cross-validation was applied to avoid sample bias in training. The ROC curve that illustrates the sensitivity and specificity of each metric is shown in Figure 3. The combined metric (solid green curve) offers better classification results than individual metrics simply because individual metrics only capture one property of a junction. At the best point, the combined metric achieves a true-positive rate of 96.3% and false-positive rate of 8%.

Figure 3.
ROC curves for junction classification. A synthetic data set of 20M 100 bp tags was generated from transcripts selected from the ASTD database. 10K true-positive junctions and 10K false-positive junctions were selected as training data sets. Five ...

We also compared the results with one of the most commonly used metrics: junction coverage (the number of tags aligned to a junction). In many studies, a junction is considered to be true if at least three tags are aligned to the junction. However, as shown in Figure 3, coverage (solid red curve) is the least reliable metric and yields the worst performance in terms of junction classification.

The specific parameters An external file that holds a picture, illustration, etc.
Object name is gkq622i198.jpg An external file that holds a picture, illustration, etc.
Object name is gkq622i199.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i200.jpg in the combined metric obtained from this synthetic data set were used for all data sets processed by MapSplice in this article. Slightly improved sensitivity can be obtained by using parameters obtained by logistic regression that are specialized for data sets with specific tag lengths. Experiments on the robustness of the parameters and their sensitivity to tag length and sampling depth are included in the Supplementary Data.

Sensitivity and specificity of splice inference

Three programs that map splice junctions using RNA-seq data were compared, MapSplice, TopHat (1.0.12) and SpliceMap (C++, v3.0, 15 April 2010). We applied all three algorithms to two representative synthetic data sets. One was a data set with 20M tags of length 50 bp. The other was a data set with 20M 100 bp tags. For both MapSplice and TopHat, we set An external file that holds a picture, illustration, etc.
Object name is gkq622i201.jpg An external file that holds a picture, illustration, etc.
Object name is gkq622i202.jpg and An external file that holds a picture, illustration, etc.
Object name is gkq622i203.jpg. For SpliceMap, the only configurable parameter is the mismatch in the segment (seeds), which was set to 1 as well. In comparison (Table 1), both TopHat and MapSplice were more memory efficient and much faster than SpliceMap. The filtering criteria adopted by SpliceMap including minimum anchor (extension) of 10 bp and no multiple alignments within a 400 kb region improved its specificity with some tradeoff in sensitivity in 100 bp tags. MapSplice performed best in both categories by detecting more true-positive junctions and fewer false-positive junctions.

Table 1.
Comparison of TopHat (21), SpliceMap (22) and MapSplice on two synthetic data sets with tags of length 50 and 100 bp, respectively

Due to the incompleteness of SpliceMap's output (tag alignments were not generated), we limit the more comprehensive comparison to TopHat and MapSplice. We investigated the sensitivity and specificity in splice inference as a function of tag length and sampling depth. We generated synthetic data sets to study the effect of these variations on junction discovery. In the synthetic data set, we have ground truth junctions and know their actual coverage, i.e. the number of tags spanning each junction. Two measures were used to evaluate the algorithms. The ‘sensitivity' is the ratio of the total number of true junctions discovered to the total number of junctions sampled in the synthetic data. The ‘specificity' is the ratio of the total number of true junctions discovered over the total number of discovered junctions. Since coverage of a junction is essential for the junction to be discovered, we plot sensitivity and specificity at coverage x as the sensitivity and specificity for all junctions with coverage x or greater, as shown in Figure 4. We also show the recovered ‘coverage ratio' for junctions identified as true in Figure 5.

Figure 4.
Sensitivity and specificity of splice inference in synthetic data sets with different characteristics. The sensitivity is the fraction of true junctions discovered among the true junctions sampled in the synthetic data. The specificity is the fraction ...
Figure 5.
Fraction of tags containing a true junction recovered (i.e. aligned to include the junction) as a function of junction coverage (defined by exponential bins). (A) TopHat recovers about 63% tags while (B) MapSplice recovers an average of 84% of the tags ...

Effect of noise

In the first experiment, we constructed the error-free and noisy versions of a 100 bp synthetic RNA-seq data sets of 20M tags as described above. MapSplice and TopHat were run on both data sets and given the same error tolerance of 4% (An external file that holds a picture, illustration, etc.
Object name is gkq622i206.jpg). Figure 5A and B shows that performance was only impaired at low coverage. When coverage is high, the sensitivity is similar despite the presence of errors. Specificity is more affected, but also converges to similar performance when coverage is high. With low coverage, more spurious junctions were discovered in the data set with error than the one without. Comparing MapSplice with TopHat, MapSplice has higher sensitivity and specificity in identifying junctions in both data sets. Specificity is substantially higher even at low coverage.

Effect of tag length

In the second experiment, we generated a synthetic data set of 20M 100 bp tags and created two additional data sets by selecting a 50 and 75 bp random subsequences of the 100 bp tags, respectively. Both MapSplice and TopHat were applied to these data sets with maximum percentage of mismatches as 4% of the tag length. The result is shown in Figure 5C and D. In general, for both TopHat and MapSplice, longer tags not only improve the sensitivity but also improve the specificity of the junction discovery. In comparison, MapSplice has higher sensitivity for all three tag lengths. The difference in sensitivity is more pronounced in junctions with low coverage, where junction discovery is most difficult.

Effect of sampling depth

In the final experiment, we generated two 100 bp data sets with a different number of tags: 10M and 20M, respectively. Doubling the sampling depth does not double the specificity of junctions, but it does improve the sensitivity. Doubling the depth has a negative effect on specificity, especially in the low coverage areas. This is mostly because increasing the number of tags sampled from a fixed number of transcripts increases the chances for a repeated tag, especially one with a high error rate, to be incorrectly aligned on the genome.

Breast cancer transcriptomes

We performed cDNA sequencing to obtain about 25 million tags of length 75 bp for four primary breast tumors and replicate samples for two breast cancer cell lines. In total, four samples correspond to the basal subtype of breast cancer, and four samples correspond to the luminal subtype. We applied both MapSplice and TopHat to detect splice junctions using the same parameter settings employed with the synthetic data sets. The mapping result is shown in Table 2. In summary, between 10% and 16% of the tags in each sample include splice junctions in their alignment. Over 77% of these canonical junctions were confirmed by known transcripts in GenBank, which represented between 6% and 11% more confirmed junctions than TopHat. MapSplice identified 2421 ~ 3173 semi-canonical junctions, much less than the number reported by TopHat. But for both sets, very similar sets of junctions are known, which suggests that MapSplice has a higher specificity for non-canonical splice junctions.

Table 2.
Tag mapping and splice junction detection results on eight breast cancer samples: two basal (BAS) primary tumors, two SUM-102 (SUM) cell lines, two luminal (LUM) primary tumors and two MCF-7 (MCF) cell lines

MapSplice reported between 1157 and 1967 non-canonical splice junctions, of which 5–8% were confirmed in known GenBank transcripts. While TopHat reported up to 5944 non-canonical junctions, none of them were confirmed in GenBank transcripts. Since the TopHat program does not search for non-canonical junctions, this result might be an artifact. We found 9205 genes that showed evidence of alternative splicing, ranging from 7371 to 8942 genes per tumor. There were 420 to 430 canonical junctions identified by MapSplice within 2 bp of a known semi-canonical or non-canonical junction. For almost all of them, the tags aligned to the canonical junction have fewer mismatches than if they were aligned to the nearby non-canonical or semi-canonical junction. Such findings suggest that there exist errors in the current database and RNA-seq data might be able to correct these errors.

MapSplice detected the expected proportion of alternative splicing categories, even though it did not rely on a database of transcript annotations. We investigated how many alternative splicing events could be detected at different minimum thresholds for the minor transcript isoforms (Table 3). For instance, at a cutoff of two or more tags per splice junction, MapSplice detected between 7535 and 8270 alternative splicing events in each tumor. These events comprised: 34.5% skipped exons; 30.3% alternative 5′-sites; 33.8% alternative 3′-sites; and 1.4% mutually exclusive exons. A previous RNA-seq study across 10 different tissues and 10 different cell lines (1) reported similar values: 35% skipped exons; 28% alternative 5′-sites and first exons; 31% alternative 3′-sites, last exons and UTRs; and 4% mutually exclusive exons. The high concordance between these two studies further suggested that the MapSplice alignments were highly accurate.

Table 3.
A survey of alternative exon splicing events identified with MapSplice junctions

We randomly selected skipped exon events for the experimental validation of MapSplice alignments to splice junctions. We calculated the proportion of splice junction tags aligning to the skipped exon isoform, and then compared this to the total number of splice junction tags aligning to either the skipped exon isoform or the included exon isoform (Figure 6). We compared these calculations with the splicing ratio determined by qRT–PCR in the MCF-7 and SUM-102 cell lines. With a Pearson's correlation of 0.84 across these 20 events, MapSplice achieved very high accuracy for splice junction counting.

Figure 6.
Correlation of exon skipping ratio detected by MapSplice and Taqman. Each point represents the exon skipping ratio measured in either the MCF-7 (black) or SUM-102 (blue) cell lines.

We identified 12 exon skipping events with significant differences between the basal and luminal subtypes. For instance, NUMB is an adaptor protein in the Notch and Hedgehog pathways with a potential skipped exon in an N-terminal PTB domain, as well as another skipped exon in a C-terminal proline-rich region (31). While all breast cancer samples had similar skipping ratios for the PTB domain exon, we detected significant differences for the skipped exon in the proline-rich region. This longer isoform had exon inclusion ratios ranging 45–78% in the luminal samples, compared with 16–22% of the basal samples (Figure 7). We expect that as more samples are sequenced, we will have more statistical power to identify alternative splicing events that can distinguish between cancer subtypes.

Figure 7.
Examples of alternative exon skipping events. The second exon in NUMB shows differential alternative splicing between two cancer subtypes. The exon skipping ratios in basal samples are ~70% while in luminal samples they are <50%.

We investigated whether molecular subtypes of tumors may have different patterns of alternative splicing regardless of their gene expression levels. We selected 129 single exon skipping events that were detected by at least three tags in each tumor. The matrix of splicing ratios was then hierarchically clustered, with each row representing a distinct splicing event and each column representing a single tumor (Figure 8). Notably, the two primary breast tumors from the luminal subtype clustered together, as did the two primary breast tumors from the basal subtype. The breast cancer cell lines clustered in between the primary tumors, which indicates that these cell lines resemble their primary tumors of origin, but also share some major differences in splicing. Principal components analysis on these splicing ratios reached similar conclusions: the first principal component distinguished cell lines from primary tumors, while the second principal component segregated luminal versus basal subtypes (Figure 8B and D).

Figure 8.
Clustering of tumor subtypes with skipping ratios of alternative exon skipping events. One hundred twenty-nine alternative exon skipping events with minimum junction support of at least three for each sample were selected. (A) Heatmap (red to blue scale) ...

DISCUSSION

Accurate identification and quantification of transcript isoforms is crucial to characterize alternative splicing among different cell types. In addition, sequence variants found within splice sites or splicing enhancer sequences may have functional consequences on alternative splicing. Thus, methods to accurately detect alternative splicing events will be necessary to determine whether these sequence variants affect the transcript isoform proportions. Since certain splice junctions can unambiguously distinguish transcript isoforms, we have focused on increasing the accuracy of aligning splice junctions de novo. For this task, we developed a new splice discovery algorithm, MapSplice, that meets three goals.

First, MapSplice performs a sensitive, complete and unbiased search to find splice junctions using approximate sequence similarity that is not dependent on features or locations of the splice sites. As a result, the algorithm can be applied equally to RNA-seq data from well-studied model organisms and also data from organisms with sparse transcripts annotations. The algorithm is capable of finding short-range as well as long-range and inter-chromosomal splices such as that might arise in gene fusion and other chimeric splicing events that result from damage to the DNA.

Second, MapSplice utilizes efficient approximate sequence alignment methods combined with a local search to create a fast and memory-efficient algorithm. Its alignment strategy can be readily generalized to reads >100 bp. With a processing power of 10 million reads (100 bp) per hour and peak memory usage below 4 GB, MapSplice can run on both desktop and servers with high efficiency.

Third, MapSplice incorporates a rigorous approach to increase specificity of the splice search, necessitated by the multiple ways in which some RNA-seq tags can find spliced alignments to the genome. By leveraging the deep sampling of the transcriptome in RNA-seq data sets, spurious splices can be discriminated from true splices. High specificity is critical as a typical RNA-seq data set can contain some evidence for hundreds of thousands of splices.

In this article, we have made a rigorous measurement of sensitivity and specificity of splice finding algorithms using realistic synthetic data sets. The performances are further assessed by experimental validation of results obtained from breast cancer samples. Using synthetic data sets, we determined that read lengths of 75 or 100 bp yield significantly better sensitivity and specificity for splice detection than 50 bp data sets. We determined that splices can be found despite the presence of errors. Finally, we used synthetic data to calibrate several filtering criteria to achieve over 98% specificity and 96% sensitivity in the detection of splice junctions in the simulated data. These filtering criteria provided superior accuracy in our comparisons to the TopHat (21) and SpliceMap (22) algorithms.

Several experimental lines of evidence also confirmed a high accuracy of the MapSplice algorithm's splice junction alignments. First, the distribution of splice junctions in various categories of alternative splicing are highly concordant with previous studies (Table 3). Second, experimental validation of 10 predictions by qRT-PCR correctly identified isoform proportions that are highly correlated (Pearson's correlation = 0.86) with their estimates based on splice junctions. Third, hierarchical clustering of splicing ratios recapitulated known molecular subtypes of four breast tumors and two breast cancer cell lines. As sample size increases, we will achieve more power to identify candidate genes with significant differences in splicing isoform proportions between molecular subtypes of cancer.

This deep sequencing study represents the first survey of alternative splicing differences between cancer subtypes. At a sequencing depth of approximately 20 million reads with a length of 75 bp, we identified between 149 722 and 178 107 canonical splice junctions, as well as 3661 to 4884 semi-canonical and non-canonical splice junctions. Notably, we discovered that 19–22% of these splice junctions have not been previously observed in full-length transcripts in GenBank. Among these junctions, ~15% connected two known exons, suggesting novel isoforms with exon skipping events.

We anticipate that tests between sample groups will be crucial to interpret data from large-scale transcriptome sequencing projects, such as the Cancer Genome Atlas. Future research efforts will be needed to distinguish splicing patterns that are enriched in a (potentially heterogeneous) disease state, compared to the natural variation in alternative splicing within populations (5).

The reconstruction of full-length transcripts from short sequence reads is a challenging task, especially for low abundance transcripts. Splice junctions constitute the building blocks for these algorithms (19,32–35). We anticipate that further advances in sequencing technologies, such as higher read depths and longer reads, will continue to improve these methods. Recent studies have combined both splice junction reads and exon reads to provide an integrated partitioning of alignments (36).

FUNDING

National Science Foundation (grant number 0850237 to J.L., J.N.M. and J.F.P.); National Institutes of Health (grant number CA143848 to C.M.P. and grant number P20RR016481 to J.L.); Alfred P. Sloan Foundation (to D.Y.C.). Funding for open access charge: National Institutes of Health (grant number CA143848).

Conflict of interest statement. None declared.

SUPPLEMENTARY DATA

Supplementary Data are available at NAR Online.

Supplementary Data:

ACKNOWLEDGEMENTS

We wish to thank Zefeng Wang, Ben Berman, Corbin Jones, Oleg Evgrafov and the anonymous reviewers for their critical comments on the manuscript.

REFERENCES

1. Wang ET, Sandberg R, Luo SJ, Khrebtukova I, Zhang L, Mayr C, Kingsmore SF, Schroth GP, Burge CB. Alternative isoform regulation in human tissue transcriptomes. Nature. 2008;456:470–476. [PMC free article] [PubMed]
2. Luco RF, Pan Q, Tominaga K, Blencowe BJ, Pereira-Smith OM, Misteli T. Regulation of alternative splicing by histone modifications. Science. 2010;327:996–1000. [PMC free article] [PubMed]
3. Andersen LB, Ballester R, Marchuk DA, Chang E, Gutmann DH, Saulino AM, Camonis J, Wigler M, Collins FS. A conserved alternative splice in the von Recklinghausen neurofibromatosis (NF1) gene produces two neurofibromin isoforms, both of which have GTPase-activating protein activity. Mol. Cell. Biol. 1993;13:487–495. [PMC free article] [PubMed]
4. Screaton GR, Bell MV, Jackson DG, Cornelis FB, Gerth U, Bell JI. Genomic structure of DNA encoding the lymphocyte homing receptor CD44 reveals at least 12 alternatively spliced exons. Proc. Natl Acad. Sci. USA. 1992;89:12160–12164. [PMC free article] [PubMed]
5. Kwan T, Benovoy D, Dias C, Gurd S, Provencher C, Beaulieu P, Hudson TJ, Sladek R, Majewski J. Genome-wide analysis of transcript isoform variation in humans. Nat. Genet. 2008;40:225–231. [PubMed]
6. Meyers GA, Day D, Goldberg R, Daentl DL, Przylepa KA, Abrams LJ, Graham JM, Jr, Feingold M, Moeschler JB, Rawnsley E, et al. FGFR2 exon IIIa and IIIc mutations in Crouzon, Jackson-Weiss, and Pfeiffer syndromes: evidence for missense changes, insertions, and a deletion due to alternative RNA splicing. Am. J. Hum. Genet. 1996;58:491–498. [PMC free article] [PubMed]
7. Pollock PM, Gartside MG, Dejeza LC, Powell MA, Mallon MA, Davies H, Mohammadi M, Futreal PA, Stratton MR, Trent JM, et al. Frequent activating FGFR2 mutations in endometrial carcinomas parallel germline mutations associated with craniosynostosis and skeletal dysplasia syndromes. Oncogene. 2007;26:7158–7162. [PMC free article] [PubMed]
8. Perou CM, Sorlie T, Eisen MB, van de Rijn M, Jeffrey SS, Rees CA, Pollack JR, Ross DT, Johnsen H, Akslen LA, et al. Molecular portraits of human breast tumours. Nature. 2000;406:747–752. [PubMed]
9. Dutt A, Salvesen HB, Chen TH, Ramos AH, Onofrio RC, Hatton C, Nicoletti R, Winckler W, Grewal R, Hanna M, et al. Drug-sensitive FGFR2 mutations in endometrial carcinoma. Proc. Natl Acad. Sci. USA. 2008;105:8713–8717. [PMC free article] [PubMed]
10. Johnson JM, Castle J, Garrett-Engele P, Kan Z, Loerch PM, Armour CD, Santos R, Schadt EE, Stoughton R, Shoemaker DD. Genome-wide survey of human alternative pre-mRNA splicing with exon junction microarrays. Science. 2003;302:2141–2144. [PubMed]
11. Pan Q, Shai O, Misquitta C, Zhang W, Saltzman AL, Mohammad N, Babak T, Siu H, Hughes TR, Morris QD, et al. Revealing global regulatory features of mammalian alternative splicing using a quantitative microarray platform. Mol. Cell. 2004;16:929–941. [PubMed]
12. Ule J, Ule A, Spencer J, Williams A, Hu JS, Cline M, Wang H, Clark T, Fraser C, Ruggiu M, et al. Nova regulates brain-specific splicing to shape the synapse. Nat. Genet. 2005;37:844–852. [PubMed]
13. Castle JC, Zhang C, Shah JK, Kulkarni AV, Kalsotra A, Cooper TA, Johnson JM. Expression of 24,426 human alternative splicing events and predicted cis regulation in 48 tissues and cell lines. Nat. Genet. 2008;40:1416–1425. [PMC free article] [PubMed]
14. Pan Q, Shai O, Lee LJ, Frey J, Blencowe BJ. Deep surveying of alternative splicing complexity in the human transcriptome by high-throughput sequencing. Nat. Genet. 2008;40:1413–1415. [PubMed]
15. Mortazavi A, Williams BA, McCue K, Schaeffer L, Wold B. Mapping and quantifying mammalian transcriptomes by RNA-Seq. Nat. Methods. 2008;5:621–628. [PubMed]
16. Sultan M, Schulz MH, Richard H, Magen A, Klingenhoff A, Scherf M, Seifert M, Borodina T, Soldatov A, Parkhomchuk D, et al. A global view of gene activity and alternative splicing by deep sequencing of the human transcriptome. Science. 2008;321:956–960. [PubMed]
17. Mereau A, Anquetil V, Cibois M, Noiret M, Primot A, Vallee A, Paillard L. Analysis of splicing patterns by pyrosequencing. Nucleic Acids Res. 2009;37:e126. [PMC free article] [PubMed]
18. Xing Y, Yu T, Wu YN, Roy M, Kim J, Lee C. An expectation-maximization algorithm for probabilistic reconstructions of full-length isoforms from splice graphs. Nucleic Acids Res. 2006;34:3150–3160. [PMC free article] [PubMed]
19. Jiang H, Wong WH. Statistical inferences for isoform expression in RNA-Seq. Bioinformatics. 2009;25:1026–1032. [PMC free article] [PubMed]
20. De Bona F, Ossowski S, Schneeberg K, Ratsch G. Optimal spliced alignments of short sequence reads. Bioinformatics. 2008;24:i174–i180. [PubMed]
21. Trapnell C, Pachter L, Salzberg SL. TopHat: discovering splice junctions with RNA-Seq. Bioinformatics. 2009;25:1105–1111. [PMC free article] [PubMed]
22. Au K, Jiang H, Lin L, Xing Y, Wong WH. Detection of splice junctions from paired-end RNA-seq data by SpliceMap. Nucleic Acids Res. 2010;2010 doi:10.1093/nar/gkq211. [PMC free article] [PubMed]
23. Langmead B, Trapnell C, Pop M, Salzberg SL. Ultrafast and memory-efficient alignment of short DNA sequences to the human genome. Genome Biol. 2009;10:R25. [PMC free article] [PubMed]
24. Li H, Durbin R. Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics. 2009;25:1754–1760. [PMC free article] [PubMed]
25. Li R, Yu C, Li Y, Lam TW, Yiu SM, Kristiansen K, Wang J. SOAP2: an improved ultrafast tool for short read alignment. Bioinformatics. 2009;25:1966–1967. [PubMed]
26. Homer N, Merriman B, Nelson SF. BFAST: an alignment tool for large scale genome resequencing. PLoS ONE. 2009;4:e7767. [PMC free article] [PubMed]
27. Li H, Ruan J, Durbin R. Mapping short DNA sequencing reads and calling variants using mapping quality scores. Genome Res. 2008;18:1851–1858. [PMC free article] [PubMed]
28. Malde K. The effect of sequence quality on sequence alignment. Bioinformatics. 2008;24:897–900. [PubMed]
29. Koscielny G, Le Texier V, Gopalakrishnan C, Kumanduri V, Riethoven JJ, Nardone F, Stanley E, Fallsehr C, Hofmann O, Kull M, et al. ASTD: the Alternative Splicing and Transcript Diversity database. Genomics. 2009;93:213–220. [PubMed]
30. Kircher M, Stenzel U, Kelso J. Improved base calling for the Illumina Genome Analyzer using machine learning strategies. Genome Biol. 2009;10:R83. [PMC free article] [PubMed]
31. Gulino A, Di Marcotullio L, Screpanti I. The multiple functions of Numb. Exp. Cell Res. 2010;316:900–906. [PubMed]
32. Heber S, Alekseyev M, Sze SH, Tang H, Pevzner PA. Splicing graphs and EST assembly problem. Bioinformatics. 2002;18(Suppl. 1):S181–S188. [PubMed]
33. Xing Y, Lee C. Reconstruction of full-length isoforms from splice graphs. Methods Mol. Biol. 2008;452:199–205. [PubMed]
34. Birol I, Jackman SD, Nielsen CB, Qian JQ, Varhol R, Stazyk G, Morin RD, Zhao Y, Hirst M, Schein JE, et al. De novo transcriptome assembly with ABySS. Bioinformatics. 2009;25:2872–2877. [PubMed]
35. Zheng S, Chen L. A hierarchical Bayesian model for comparing transcriptomes at the individual transcript isoform level. Nucleic Acids Res. 2009;37:e75. [PMC free article] [PubMed]
36. Richard H, Schulz MH, Sultan M, Nurnberger A, Schrinner S, Balzereit D, Dagand E, Rasche A, Lehrach H, Vingron M, et al. Prediction of alternative isoforms from exon expression levels in RNA-Seq experiments. Nucleic Acids Res. 2010;8:e112. [PMC free article] [PubMed]

Articles from Nucleic Acids Research are provided here courtesy of Oxford University Press
PubReader format: click here to try

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...

Links

  • MedGen
    MedGen
    Related information in MedGen
  • PubMed
    PubMed
    PubMed citations for these articles

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...