What Pca Analysis Tells Us About Performance Of Different Gene Regulatory Network Inference Methods

Question

[jack]

Hi all,

I a read paper about gene regulatory network inference algorithms. they reconstructed GRN with different GRN inference methods and they used first two principal components to see the clusters. Can somebody explain me, what PCA say us about performance of these algorithms ? for me seems that, if different methods produce similar GRN, then the PCA of them should reveal same number and similar clusters. is it right?

[Devon Ryan]

What did they use the principal components of? The principal component of a network doesn't make any sense in and of itself. I can only assume that they did PCA on the expression data (rather than networks) to see if the genes group into some coherent number of obvious clusters. BTW, this seems like a pretty hand-wavy way to go about assessing performance.

[jack]

they reconstructed GRN with regression method, Bayesian and mutual information methods. then they used PCA on inferred GRNs by these methods to see how similar the networks are .

[Woa]

Can you cite the paper?

Answer 1

[Charles Warden]

The PCA results will depend upon the genes taken into consideration.

For example, consider a PCA plot for all genes on a micoarray or RNA-Seq experiment versus a list of ~1000 differential expressed genes (or ~1000 genes in a regulatory network). There will probably be less variation and tighter clustering in the filtered list.

On the other hand, if you asked how the PCA plot for a list of genes that varied with a |fold-change| > 1.5 and FDR < 0.05 compared to a plot of expression from several very similar GRNs, there probably won't be much difference (if the network construction is doing something useful), especially if you are dealing with something like a homogeneous cell line experiment where clustering is the same for both the genome-wide and filtered gene lists.