STAR-Fusion is a program that detects RNA fusions events in RNA-Seq data. According to the paper describing the program, STAR-Fusion is much better than the dozen or so other callers under active development.
Still, reading the paper left me with some questions. As descried by the authors, STAR-Fusion is not just a good caller, but the best caller by a wide margin. See Fig 3A. The next nine best callers have AUC values of 0.5 to 0.3, but STAR-Fusion has a value of 0.8 in the author’s testing.
And what is the source of this incredible result? The authors are silent on the subject. They don’t know, or perhaps didn’t notice how remarkable their achievement is, and so don’t remark on it. The description of the STAR-Fusion algorithm seems very similar to the algorithms used by every other RNA fusion caller. Some do better than others, so details of implementation must matter.
So what is the critical advance STAR-Fusion makes? Is better sequence alignment key? Is the filtering approach? The paralog handling seems like it cuts down on false positives, is this key? Discovering the critical factors for RNA fusion calling would be an important result.
Or are the performance results in the paper dependent on the synthetic test data set the authors use? Will subsequent papers comparing STAR-Fusion to other methods find that it is only average, or sub-par?