Benchmarking short-read germline structural variant calling highlights advantages of using ensembles of tools and small impact of graph genome alignment

Benchmark of 14 short-read germline SV callers and ensemble strategies on a pedigree-based truth set

The authors evaluated 14 short-read germline structural variant (SV) callers on four samples from a recently published, high-confidence pedigree-based truth set, together with an in-house dataset; its recent release reduces the risk that it was used to train the tools. DRAGEN achieved the best overall performance (mean recall 0.36, mean F1 0.51); among open-source tools, Manta and Dysgu had the highest recall (0.26 each) and Manta the highest F1 (0.41). Several multi-caller ensemble strategies reached, in some settings, higher recall than any individual tool, with Dysgu, Octopus, Manta and Tardis among the top-performing sets. Frequency filtering reduced the downstream interpretation load by nearly half, and alignment to a graph genome assembly had only a small effect compared with a linear reference.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The figure to remember is the ceiling: even the best tool recovers on average only about a third of the truth-set SVs, a reminder that a negative short-read genome does not rule out a causal SV. The gain from ensemble strategies is reported only "in some settings", and the abstract states neither the size of the in-house dataset nor code availability. Frequency filtering, which nearly halves the number of variants to review, is the result most directly transferable to a diagnostic pipeline.

Analysis by Dr Thibaut Benquey