Fig. 4

a Distributions of properties of TP, FP, and FN RIs across short-read detection tools. Each panel displays the boxplot distributions of intron length (top, log scale), % of bases in the intron that are G or C (2nd row), relative position in transcript (3rd row), and % of intron bases with overlapping annotated exons (bottom) for the output from each of eight short-read tools (from left to right: IRFinder-S, superintronic, iREAD, KMA, IntEREst, rMATS, MAJIQ, and SUPPA2). Y-axes correspond to intron properties as labeled, with each boxplot along the x-axis corresponding to the TP (green, left boxes), FP (pink, middle boxes), and FN (blue, right boxes) calls for HX1 (left) and iPSC (right). b Short-read tool performance as a function of intron length. Each panel depicts the LOESS-smoothed precision (top), recall (middle), or F1-score (bottom) in either the HX1 (left) or iPSC (right) sample across overlapping, sliding window intron length ranges (Section 5). Smooths are grouped and colored by eight short-read tools (red = IRFinder-S, yellow = superintronic, green = iREAD, purple = IntEREst, blue = KMA, light pink = MAJIQ, light green = rMATS, light blue = SUPPA2). c Read coverage and exon overlap as a function of position within an intron. LOESS-smoothed short-read data (see Section 5) show the median log10-scaled coverage (top row, y-axes) and fractions of introns with overlapping exons (bottom row, y-axes) as a function of position (x-axis, 5′ \(\rightarrow\) 3′ on positive strand) for HX1 (left column) and iPSC (right column). Introns were grouped by truth category membership for at least 3/8 tools (colors, blue = FN, pink = FP, green = TP)