A good class separation metric was given by Slonim et al.:
P(g,c) = (u_{1} - u_{2})/(s_{1} + s_{2})
where
- c: a class vector
- g: an expression vector of a gene over n samples
- u_{1}: mean expression level in class 1
- u_{2}: mean expression level in class 2
- s_{1}: standard deviation in class 1
- s_{2}: standard deviation in class 2