I'm having problem interpreting the output of perf(PLS), and I could not find any case to relate to in the MixOmics documentation.
My dataset is quite small, I want to relate a 12x160 matrix (expression values for 160 immune genes in 12 samples) with a 12x11 matrix (viral loads for 11 viruses in the same 12 samples). I ran a PLS with 4 and 9 components with the following commands:
PLS <- pls(Y=PLS_data$genes,X=PLS_data$pathogens,ncomp=4, mode = "regression") (or ncomp=9) tune.PLS <- perf(PLS, validation = "Mfold", folds = 10, progressBar = FALSE, nrepeat = 100) plot(tune.PLS$Q2.total)
The output perf_plot.pdf is attached. I'm confused about what to take away from it. I would have thought that the first few relevant Q2 values would be above the 0.0975 threshold, but when I run it with 4 components the first one as well as the third and fourth one are above the threshold, and not component 2. Similar story when I run a PLS with 9 components, the components above the threshold are not always the first ones (see attached plot).
First, I have to admit I'm still not exactly sure what these Q2 values are. I thought it was comparable to the variance explained by principle components (similar to PCAs) but the fact that the Q2 values are not always decreasing with increasing number of latent variables doesn't make sense to me. But accepting that, I should only retain for subsequent analysis like Clustered Image Maps components 1, 4 and maybe 3. But then what to make up of the ninth component when running a PLS with 9 components? Preliminary PCAs on both viruses and immune genes dataset indicate that up to 4 principal components are sufficient to explain most of the variation.
