Mikado: Leveraging Multiple Transcriptome Assembly Methods for Improved Gene Structure Annotation

The reconstruction of transcript sequences from RNA-Seq reads is a powerful approach to study alternative splicing and improve gene structure annotation. However, while numerous methods have been proposed to infer the original transcript structure from sequencing data, all of them exhibit idiosyncratic strengths and weaknesses that make them particularly adept at reconstructing certain classes of transcripts while failing with others. The best performing methods identify ~21% of spliced protein coding transcripts accurately in human with substantial variation in accuracy across methods. Pooling together the assemblies of multiple methods provides a potential strategy to improve sensitivity, but creates the problem of determining which of the alternative assemblies best represents the correct structure at each locus. To solve this conundrum, we have developed a novel algorithm capable of defining expressed loci and selecting the best transcript assemblies at each position. We tested our method on human, D. melanogaster and A. thaliana, recovering several thousand gene structures that were either missed entirely, incorrectly fused or with erroneous or incomplete gene models in the best of the assembly methods tested, while also removing the majority of incorrect/undesirable gene structures. The algorithm is implemented in a novel software tool, Mikado, currently under active development.