Misinterpretation of RNA-Seq Data Widespread

In a paper published today in PLOS Biology, Tel Aviv University scientists point out a frequent technical bias in data generated by RNA-seq technology that recurrently leads to false results and also show how this bias can be removed from the data.

Analyzing dozens of publicly available RNA-seq datasets, which profiled the cellular responses to numerous different stresses, lead author Shir Mandelbaum and colleagues noticed that sets of particularly short or long genes repeatedly showed changes in expression level.

Puzzled by this recurring pattern, the authors then asked whether it reflects some universal biological response common to many different triggers or it rather stems from some experimental artefact. To tackle this question, they compared replicate samples from the same biological condition. Differences in gene expression between replicates can reflect technical effects that are not related to the experiment's biological factor of interest. Unexpectedly, the same pattern of particularly short or long genes showing changes in expression level was observed in these comparisons between replicates, demonstrating that this pattern is the result of a technical bias that seemed to be coupled with gene length.

A main goal of RNA-seq experiments is to characterize biological processes that are activated or repressed in response to the conditions of interest. Notably, specific biological processes are executed by products of particularly short and long genes. For example, many of the short genes encode proteins that constitute the ribosome. Conversely, many of the long genes encode proteins that constitute the extra-cellular matrix (ECM.

Mandelbaum and colleagues were able to show how, in many RNA-seq datasets, the length bias they detected, combined with some flaws in the statistical analysis, can lead to the false identification of specific biological functions (including ribosome and ECM-related functions) as cellular responses to the conditions tested. Importantly, the study also shows how this bias can be removed from the data, thus filtering out false calls while preserving the biologically genuine ones.