Comparison of gene fusion detection algorithms reveals frequently overlooked driver fusions in hematologic malignancies

Accurate detection of driver gene fusions is essential for the diagnosis, treatment, and prognostic prediction of hematologic malignancies. Despite increasing reliance on RNA sequencing (RNA-seq) for fusion detection, its algorithms have not been sufficiently examined for their ability to detect clinically relevant driver fusions. We evaluated 12 algorithms using conventional RNA-seq from 170 cell lines and targeted RNA-seq from 26 cell lines and 165 clinical samples. The true positive rate, based on 61 and 24 driver fusion-cell line pairs for conventional and targeted RNA-seq, varied between 0.41-1 (median 0.81) and 0-1 (median 0.85), respectively. Many algorithms failed to detect fusions resulting from small deletions (including STIL::TAL1 and FIP1L1::PDGFRA), lowly expressed fusions, and IGH fusions (DUX4::IGH and IGH::NSD2). Targeted RNA-seq more sensitively detected driver fusions than conventional RNA-seq, especially lowly expressed ones. One algorithm, Arriba, detected all driver fusions. These findings will inform algorithm selection in clinical settings.