Open Access Highly Accessed Open Badges Research article

Blind spots of quantitative RNA-seq: the limits for assessing abundance, differential expression, and isoform switching

Hubert Rehrauer1*, Lennart Opitz1, Ge Tan12, Lina Sieverling1 and Ralph Schlapbach1

Author Affiliations

1 Functional Genomics Center Zurich, University of Zurich/ETH, Zurich, Switzerland

2 Clinical Sciences Centre, London W12 0NN, United Kingdom

For all author emails, please log on.

BMC Bioinformatics 2013, 14:370  doi:10.1186/1471-2105-14-370

Published: 24 December 2013



RNA-seq is now widely used to quantitatively assess gene expression, expression differences and isoform switching, and promises to deliver results for the entire transcriptome. However, whether the transcriptional state of a gene can be captured accurately depends critically on library preparation, read alignment, expression estimation and the tests for differential expression and isoform switching. There are comparisons available for the individual steps but there is not yet a systematic investigation which specific genes are impacted by biases throughout the entire analysis workflow. It is especially unclear whether for a given gene, with current methods and protocols, expression changes and isoform switches can be detected.


For the human genes, we report their detectability under various conditions using different approaches. Overall, we find that the input material has the biggest influence and may, depending on the protocol and RNA degradation, exhibit already strong length-dependent over- and underrepresentation of transcripts. The alignment step aligns for 50% of the isoforms up to 99% of the reads correctly; only in the presence of transcript modifications mainly short isoforms will have a low alignment rate. In our dataset, we found that, depending on the aligner and the input material used, the expression estimation of up to 93% of the genes being accurate within a factor of two; with the deviations being due to ambiguous alignments. Detection of differential expression using a negative-binomial count model works reliably for our simulated data but is dependent on the count accuracy. Interestingly, using the fold-change instead of the p-value as a score for differential expression yields the same performance in the situation of three replicates and the true change being two-fold. Isoform switching is harder to detect and for at least 109 genes the isoform differences evade detection independent of the method used.


RNA-seq is a reliable tool but the repetitive nature of the human genome makes the origin of the reads ambiguous and limits the detectability for certain genes. RNA-seq does not equally well represent isoforms independent of their size which may range from ~200nt to ~100′000nt. Researchers are advised to verify that their target genes do not have extreme properties with respect to repeated regions, GC content, and isoform length and complexity.