If we consider a list of genes and their associated expression levels (fold changes), one would most probably tempted to introduce as noise variable the mean or the median (considering that the expression levels are part of a gaussian distribution).
What other parameters should be considered when trying to build a more accurate model ?