Value Specific Variance
In practice, we have found that the variance
of expression measurements depends on the magnitude of expression
values or fold changes. In the figure below we present the variance
associated with different log fold change values. This variance was
computed from the 2 repeats in the fkh1/2 dataset by placing a Gaussian
bump around the log fold changes, and using the difference between
the two repeats weighted by their distance from the value of interest.
The complete details of this procedure are described in the
following writeup. As
can be seen, the higher the absolute log fold change, the higher the
associated variance. Since most values are relatively low,
using the same variance for all values results in an under estimate of the
variance for genes that have large changes during the cell cycle.
This results in a large number of false positives since small
shifts in the magnitude of expression values can result in large
global differences for genes with high fold change values. Instead,
we incorporate the value specific variance into our algorithm, as
discussed in the paper.
