Fig. 4
From: Magic-BLAST, an accurate RNA-seq aligner for long and short reads
ROC curves showing intron discovery as a function of minimal read coverage, from 1 to 100 (or to the maximal observed coverage). The point with coverage 1, corresponding to all introns found in the experiment, is the point closest to the top right corner. The plots show, for each minimal coverage, the true positives on the Y-axis and the false positive on the X-axis. In the experimental sets, annotated and unannotated introns are used as a proxy for true and false positives. The best curve would have all the true positives before any false positives, meaning the steeper the slope the better. The benchmark sets, iRefSeq and six Baruzzo, have a built-in truth (vertical blue line in some graphs). a For the iRefSeq set, because of the alternative splice variants, the truth has introns supported by 1, 2, and up to 51 RefSeqs, and Magic-BLAST (red) follows the truth remarkably closely, point by point. It finds slightly more true-positive introns than the HISAT2 programs, but the biggest difference is that HISAT2 finds fifteen to seventeen times as many false positive introns. STAR long finds only 60% of the introns with some false positives. b For PacBio, with less than 9000 reads, Magic-BLAST already finds 11464 annotated introns, many more than the two HISAT2 versions and STAR long, yet it finds the fewest unannotated introns. c The Roche presents a similar, though less extreme, result. d In Illumina (zoomed in Additional file 1: Figure S3.1), Magic-BLAST followed by STAR 1-pass have the steepest slopes. Then come STAR 2-pass and TopHat2, then HISAT2; these last three aligners call unannotated introns at high coverage. e-j) In the Baruzzo shallow human T1 (e) and T2 (g) benchmarks, Magic-BLAST then HISAT2 perform the best, followed by STAR 1- then 2-pass, then TopHat2. i In human T3, HISAT2 and TopHat2 drop considerably, only Magic-BLAST and STAR can find introns in the presence of a high level of mismatches. f, h, j In the ultra-deep malaria sets, Magic-BLAST remains best, STAR 2-pass and HISAT2 drop below TopHat2. At coverage 1, STAR has by far the largest number of false positives (Additional file 1: Figures S3)