Three principal approaches have been proposed for inferring the set of transcripts expressed in RNA samples using RNA-seq. The simplest approach uses curated annotations, which assumes the transcripts in a sample are a subset of the transcripts listed in a curated database. A more ambitious method involves aligning reads to a reference genome and using the alignments to infer the transcript structures, possibly with the aid of a curated transcript database. The most challenging approach is to assemble reads into putative transcripts de novo without the aid of reference data.
Researchers at the University of Cambridge and Imperial College London have systematically assessed the properties of these three approaches through a simulation study. They have found that the sensitivity of computational transcript set estimation is severely limited. Computational approaches (both genome-guided and de novo assembly) produce a large number of artefacts, which are assigned large expression estimates and absorb a substantial proportion of the signal when performing expression analysis. The approach using curated annotations shows good expression correlation even when the annotations are incomplete. Furthermore, any incorrect transcripts present in a curated set do not absorb much signal, so it is preferable to have a curation set with high sensitivity than high precision.
Software to simulate transcript sets, expression values and sequence reads under a wider range of parameter values and to compare sensitivity, precision and signal-to-noise ratios of different methods is freely available online (https://github.com/boboppie/RSSS) and can be expanded by interested parties to include methods other than the exemplars presented in this article.