Regarding 3D PCA Plot in R

Question

[Nitin]

Hello,

I have 4 different microarray datasets from different samples (A(6 replicates),B(10 replicates),C(10 replicates),D(4 replicates)). In order visualize how these samples cluster together I performed PCA in R and generated 3dplot (3 PC components) (Image is attached )with 3dgrid. This shows that A and C are clustered separately. But B and D are together in this plot I see that B cluster is above the grid and D is below the grid. I understand they are together my question is how to interpret this cluster (B and D) closer look at this plot shows that B points are above the grid and D is below the grid what is the meaning of this?

Thanks,
Sa

[Giovanni M Dall'Olio]

To me it seems a bit weird to do a 3D plot of a PCA analysis. The motivation of doing a PCA is to reduce a multi-dimensional dataset to a lower number of dimensions, by 'rotating' the multi-dimensional plane until you can see the best separation between the samples. However if you represent the PCA results as a 3-D plane and start rotating it, you are kind of doing a manual PCA on the PCA results itself. At this point it will be quite difficult to interpret the results.

[Irsan]

It's even more easier to judge on how the samples cluster when you do unsupervised hierarchical clustering.

[Martombo]

I think it's usually easier to look at a PCA plot if you use only two dimensions. If you really want to look at the third component, you can plot it against the second for example.

Answer 1

[Devon Ryan]

The sign of a value (i.e., + or -) in a PCA has no actual meaning. Your groups generally cluster separately, that's always nice. Presumably when you do the real statistics on the samples you'll find a number of robust findings, then (though not seeing group clustering in a PCA plot doesn't preclude this).

[Nitin]

Hi Devon,

Thanks for clarification so plot says that B and D are clustered separately right?

Let me know,

Thanks, Sai

[Devon Ryan]

Yes, they're separated by PC3. Note that it's good to look at how much variation is actually accounted for by each of the principal components you plot. Also, you might find it easier to look at 3 2D plots, rather than a 3D one.