Hi all, I need to do some quality control on a two channel Agilent dataset, of which unfortunately I can't retrieve the raw data. The values are in log ratios and when I do a PCA I can't see any meaningful clustering but rather all the points are clumped together. Are the log ratios causing that or it just shows there isn't much going on biologically? if the problem are the log ratios is there a sensible way to transform them or carry out other kind of PCAs? Thanks a lot

In principle, there's nothing wrong in doing PCA on log ratio microarray data unless there's something about the data that you're not telling us. You could try multi-dimensional scaling (but note that PCA and classical/metric MDS are the same when using Euclidean distance).

So you are doing PCA on a matrix of probes and logFC, where each logFC column comes from a two-colour array. I'm not sure in this case PCA can give anything meaningful since it depends on what samples have been paired, unless one channel is always a reference common to all the arrays.

It might be possible to extract the individual intensities if you have the intensity value of each spot besides the logFC value.

I have a matrix file with only the log ratio and no reference which i downloaded from GEO (Ac. Number: GSE50988). I did contact the author for the raw data but I didn't get any reply. This is the PCA plot

in theory samples should cluster at least to some degree by time points, but here it looks quite messy. It could be that the data is very noisy, i would just like to know if there is a way i can tell that before going to do downstream analyses. I will try with the multi-dimensional scaling, any other suggestions will be much appreciated!

Its normal for the samples to look this way. The pca uses all probes and the vast majority are consistent between samples. What you're seeing is the dominant signal is unchanging between samples, probably they all look like the same tissue type.the differences are hidden because they're only present at a minority of probes.

In principle, there's nothing wrong in doing PCA on log ratio microarray data unless there's something about the data that you're not telling us. You could try multi-dimensional scaling (but note that PCA and classical/metric MDS are the same when using Euclidean distance).