Affymetrix Normalization And Non-Linear Relationships

Question

[tonja.r]

I am reading the paper of Bolstad 'A comparison of normalization methods for high xensity oligonucleotide array data based on variance and bias' and got stuck at the line

Affymetrix has approached the normalization problem by proposing that intensities should be scaled so that each array has the same average value. The affymetrix normalization is performed on expression summary values. This approach does not deal particularly well with cases wrehe there are non-linear relationships between arays.

I am not quite sure I undestand why this approach does not preform good with non-linear relationships.

Thanks in advance

Answer 1

[Devon Ryan]

Suppose you have a dataset like:

a <- c(1:100) #the first array
b <- c(1:90,201:210) #the second array
plot(a/mean(a), b/mean(b))
abline(a=0, b=1)

One would expect the points to cluster around the line, but they don't in this case. If one had instead used a method that incorporates quantile normalization (e.g., RMA), then the results would have been better (a bit less so in this case, but still better than just dividing by the average).

[pld]

Even then does quantile normalization fair that well? I thought that the problem of intensity mean driving intensity variance made quantile normalization behave poorly in the upper tail of intensities.

[Devon Ryan]

Yeah, in hindsight my example was probably not that great.