Statistical assessment of gene fusion detection algorithms using RNA Sequencing Data

The detection and quantification of fusion transcripts has both biological and clinical implications. RNA sequencing provides a means for unbiased and high resolution characterization of fusion transcript information in tissue samples. We evaluated two gene fusion detection algorithms, TopHat-Fusion and deFuse, using transcriptomes derived from breast cancer biopsies and cell line. The two algorithms use different strategies to reduce false positive fusion calls: TopHat-Fusion uses a cascade of heuristically designed filters while deFuse relies on statistical methodologies to assess and prune potential fusions. We evaluated the concordance of gene fusion calls between the two algorithms and characterized the evidence generated by each methodology for the common fusions. Our results show that the concordance between the two algorithms is about 6% at comparable number of fusions detected. deFuse tends to be more sensitive in detecting fusions when compared to TopHat-Fusion. On average, deFuse identifies proportionally more reads than TopHat-Fusion as evidence for concordant fusions. While deFuse provides a more quantitative summary of the reliability of the fusion calls, TopHat-Fusion is able to detect additional fusion transcripts that were missed by deFuse. Further research is required to develop algorithms that can detect fusions with high sensitivity and specificity.