24 months ago by
Cambridge, United Kingdom
No, there is no general objective justification for any particular log-fold change threshold. Mathematically speaking, it is possible to reject the null hypothesis at any non-zero log-fold change if the variability is low enough. One could argue that small log-fold changes are not biologically relevant, but the exact definition of "small" is open to interpretation. Larger log-fold changes are also more robustly detected across technologies (e.g., RNA-seq and qPCR), though selecting a threshold on this basis would depend on the sensitivities of the technologies involved. Somewhere between 1.1 to 1.5 is a common choice for a "sensible" threshold.
But all this is getting away from the main point, which is the detection of DE genes. If you want to do this in a statistically rigorous manner, use the BH-adjusted p-values to control the false discovery rate. This ensures that the expected proportion of false positives in your set of significant DE genes is below a certain threshold (usually 5%). Now, you might say that this approach also involves the selection of an arbitrary threshold. However, with this approach, at least the choice of threshold is directly related to the probability of whether the genes are truly DE or not. A log-fold change threshold doesn't tell you much about the error rate, as it doesn't account for the variability of the expression values.
Finally, if you do need a log-fold change threshold, the
treat function should be used, and DE genes selected on the basis of the adjusted p-values. This ensures that the FDR is controlled while only considering genes with log-fold changes above a minimum value.
modified 24 months ago
24 months ago by
Aaron Lun • 24k