# Selecting genes with more contribution from PCA

I have RNA-seq data in response to treatment vs non response; By machine learning I selected three principle components likely can predict the response based on the gene expression. Now I have something like this

> head(pca$$rotation[1:10,1:6]) PC1 PC2 PC3 PC4 PC5 PC6 ACTB -0.01858031 -0.005810636 0.014939832 -0.009594126 -0.0118313644 -0.011534401 ATP5F1 -0.01822611 0.014851900 -0.016401920 -0.015616379 0.0254850111 -0.010225516 DDX5 -0.02477827 -0.020573978 0.002199966 -0.009296852 0.0006222345 0.014506614 EEF1G -0.02803284 0.001126758 -0.010570521 -0.017035205 0.0007583815 0.011225639 GAPDH -0.02691309 0.005503228 0.006446502 -0.012686738 -0.0289737909 0.007036631 NCL -0.03019765 -0.003011778 -0.015534611 -0.002019556 -0.0126401245 0.013977142 > > dim(pca$$rotation)
[1] 3962   56
>


Machine learning says PC 6, PC 21 and PC 39 are better; how I know from 3962 genes in each of these PCs which genes is more contribute to response to treatment?

• Have you looked at the loadings in those PCs? – Devon Ryan Oct 4 '19 at 9:43
• Hi @Fereshteh you should know this, you have enough experience in ML. PCA cannot solve this but you can use the classification for supervised learning which will tell you – Michael G. Oct 4 '19 at 9:45
• Like @DevonRyan mentioned, you can go by the loadings, which will be sort( abs(pca\$rotation[,"PC6"]),decreasing=T) for example in your case. But it kind of defeats the purpose of PCA right? You use PCA to get linear combinations of the genes.. – StupidWolf Oct 4 '19 at 9:55
• Yes the genes with the higher loadings in the better PCs should contribute more towards response / non-response. My point is, it should be a combination of the genes. For example, if PC6 has strong +ve loadings for gene 1,2,3 and strong -ve loadings for gene 4,5,6, I would interpret it as, a strong expression of genes 1-3 and low expression of genes 4-6 contribute towards a response phenotype – StupidWolf Oct 4 '19 at 11:31
• What exactly is the goal of your analysis? If it is to predict and then point out a few potential causative genes, then the PCA approach you have is ok. I would be cautious about using up to PC39 (which explains very little of the variance in gene expression). – StupidWolf Oct 4 '19 at 11:33