Fig. 5

The figure shows the accuracy rate as a function of the recall rate for Gridss [30] (orange), Manta [3] (green), and our approach (blue) on the two yeast genomes YPS138 (as reference genome) and UFRJ50816 (as sequenced genome). Here the accuracy rate is the percentage number of correct entries among all reported entries. A reported entry is considered as correct if it matches the position of a ground truth entry. The recall rate is the number of correctly reported entries over the number of entries in the respective ground truth matrix. A shows benchmarking for simulated reads. There, we simulate 100 × coverage for 250-nt-long Illumina reads with DWGSIM [36] using default parameters, while setting the mutation rate to zero. Furthermore, we simulate 100 × coverage for CCS-PacBio reads using SURVIVOR [37] using default parameters. B displays benchmarking results for the original Illumina HiSeq 2500 and PacBio SMRT reads that were utilized for the assembly of UFRJ50816 in [23]. The solid curves (\(\pm\) 0nt) and triangles benchmark the SV callers for their ability to rediscover the exact locations of breakends. The dotted curves (\(\pm\) 25nt) and discs show the callers’ performance if we grant a tolerance of \(\pm\) 25nt for the breakends positions. Here, if an SV caller reports multiple entries within the emerging 50nt window, we pick the entry with the highest score and discard all remaining entries. Manta does not report different confidences and hence appears as single points in the analysis. For all other callers, curves are inferred by gradually adding calls to the analyzed set, in reverse order of their confidence. Gridss and Manta are designed to work with short reads (Illumina reads) merely and are therefore excluded from the PacBio benchmarking. The reported VCF BND-tags of Manta and Gridss are translated into corresponding matrix entries for benchmarking