OLego: fast and sensitive mapping of spliced mRNA-Seq reads using small seeds

Abstract

A crucial step in analyzing mRNA-Seq data is to accurately and efficiently map hundreds of millions of reads to the reference genome and exon junctions. Here we present OLego, an algorithm specifically designed for de novo mapping of spliced mRNA-Seq reads. OLego adopts a multiple-seed-and-extend scheme, and does not rely on a separate external aligner. It achieves high sensitivity of junction detection by strategic searches with small seeds (~14 nt for mammalian genomes). To improve accuracy and resolve ambiguous mapping at junctions, OLego uses a built-in statistical model to score exon junctions by splice-site strength and intron size. Burrows-Wheeler transform is used in multiple steps of the algorithm to efficiently map seeds, locate junctions and identify small exons. OLego is implemented in C++ with fully multithreaded execution, and allows fast processing of large-scale data. We systematically evaluated the performance of OLego in comparison with published tools using both simulated and real data. OLego demonstrated better sensitivity, higher or comparable accuracy and substantially improved speed. OLego also identified hundreds of novel micro-exons (<30 nt) in the mouse transcriptome, many of which are phylogenetically conserved and can be validated experimentally in vivo. OLego is freely available at http://zhanglab.c2b2.columbia.edu/index.php/OLego.

Figure 1.

Overview of OLego. Each read is processed independently by OLego. (1) Continuous mapping to the genome or exonic alignment is attempted first. If no hits are found within the allowed number of mismatches, junction alignment is searched through steps starting from (2) seeding (3) seed mapping and hit clustering into candidate alignments, and (4) candidate-exon identification and extension. (5) Junctions are then searched between two consecutive candidate exons and at the end of the read, and small exons are searched when necessary. (6) Finally, exons and junctions are connected and ranked to identify the optimal alignment for the whole read.

Figure 2.

Sensitivity of junction detection at different coverages. (A) Tests on 10 million 2 × 100-nt simulated reads; (B) Tests on 10 million 2 × 150-nt simulated reads. For each panel, the simulated junctions were binned according to their coverage, from 1 read per junction to >4 reads per junction. The true numbers of junctions in the simulation are shown by lines with markers on the right axis, and the sensitivity of OLego, MapSplice, TopHat and PASSion are indicated by bars on the left axis.

Figure 3.

Comparison of mapping speed. (A) Tests on 2 × 100-nt simulated reads; (B) Tests on 2 × 150-nt simulated reads. Running time (wall time) of TopHat (square) and OLego (triangle) on 10 million simulated paired-end reads with different numbers of CPU cores is shown. The values were averaged across three replicates for each test, with error bars indicating standard deviations.

Figure 4.

Discovery of small and micro-exons in simulated mRNA-Seq data. (A) 2 × 100-nt simulated reads; (B) 2 × 150-nt simulated reads. In each panel, internal exons within mapped reads were counted. The numbers of true (open columns) and false (solid columns) exons of different sizes, compared with the ground truth (horizontal bar) are shown for OLego, TopHat, MapSplice and PASSion, respectively. The overall sensitivity (SN_9–39) and the sensitivity for exons of size 9–15 nt (SN_9–15) are indicated on each plot.

Figure 5.

Distributions of exon junctions discovered in mouse retina mRNA-Seq data. (A) The junctions found by OLego were binned according to the numbers of supporting reads. Different patterns indicate categories of junctions in the bar plot: annotated junctions; junctions with both splice sites annotated (Class I novel); junctions with only one splice site annotated (Class II novel) and junctions without any splice site annotation (Class III novel). The total number of junctions discovered in each bin is shown by the solid line with axis on the right. (B) The junction alignments were grouped according to their anchor sizes. The categories of the junctions are shown in the same way as in panel (A), and the numbers of junction alignments are shown by the solid line with the y-axis on the right.

Figure 6.

Discovery of micro-exons in mouse retina mRNA-Seq data. (A) Number of micro-exons identified by OLego. Exons are binned by their sizes (∼9–27 nt), and in each bin, they are classified into three groups: annotated micro-exons in previous gene models (black), high-confidence novel micro-exons (exons with both flanking constitutive splice sites annotated; gray) and other (blank). (B) Cumulative distribution of exon inclusion level for annotated and high-confidence novel micro-exons; only those cassette exons with ≥10 reads that support either isoform were included for this analysis. (C) The distribution of total splice-site score (3′ + 5′ splice sites) for each group of micro-exons is shown as a boxplot. (D) The pyrimidine (C/U) content in the upstream 100-nt intronic sequences, calculated using 10-nt sliding windows. (E) Cross-species conservation around the micro-exons. The medians of phastCons scores across 30 vertebrate species in the intronic regions immediately upstream and downstream of the annotated and high-confidence novel micro-exons are shown. (F) An example of a 9-nt novel micro-exon in the Kcnn2 gene is shown. This exon is missing in current gene models (e.g. RefSeq) or cDNA/EST data (not shown), but both isoforms are abundant in the mouse retina (the two tracks on the top). The micro-exon is embedded in a longer stretch of conserved sequences.

Figure 7.

Experimental in vivo validation of micro-exons discovered by OLego. (A, C) Primers were designed either in the flanking exons to detect both micro-exon inclusion and skipping isoforms (A), or at the exon junction to specifically detect micro-exon expression (C). Primers positions and structure of each isoform are indicated (not to scale). (B, D) RT-PCR analysis of micro-exon expression in mouse retina using primers described in (A, C). Micro-exon included and skipped isoforms are indicated next to the corresponding bands by solid and empty arrowheads, respectively. (E) Correlation of micro-exon inclusion ratios estimated from mRNA-Seq data and those measured by radioactive PCR, as described in (B) (n = 3).