Difference in PCA variance calculation in DESeq2
1
0
Entering edit mode
ag1805x ▴ 50
@ag1805x-15215
Last seen 20 days ago

I was trying to plot PCA using DESeq2 plotPCA function and prcomp function. However, the variances I obtained was quite different. Why is this?  Code for PCA using prcomp:

pca <- prcomp(t(countsPC_batch))
percentage <- round(((pca$sdev^2) / (sum(pca$sdev^2))) * 100, 2)
pca_data <- data.frame(pca$x, SampleType=factors_new$SampleType, StudyAccession=factors_new\$StudyAccession)
tiff(filename=paste0("Sample_PCA", OutputNumber, ".tiff"), height=10, width=10, units='in', res=300)
ggplot(pca_data,aes(x=PC1,y=PC2, shape=SampleType, col=StudyAccession )) +
geom_point(size = 4) +
labs(title="Sample PCA", subtitle=paste0("Samples = ", SamplesUsed, " Normalization=", NormalizationUsed))+
xlab(paste0("PC1: ", percentage, "% variance")) +
ylab(paste0("PC2: ", percentage, "% variance")) +
theme(...)
dev.off()


The Proportion of Variance from summary(pca) was consistent to the calculated percentages.

Further, through hierarchical clustering, I observed two major clusters, but in these PCA I think there are three groups.

deseq2 pca hclust • 542 views
0
Entering edit mode
@mikelove
Last seen 3 days ago
United States

Take a look at ?plotPCA which I think will answer your question.

0
Entering edit mode

Thank you Mike. So it performs PCA on the top 500 genes by variance.

Can you help me with the second part of the question:

Further, through hierarchical clustering, I observed two major clusters, but in these PCA I think there are three groups.

1
Entering edit mode

Sure, these are just different techniques at visualizing high dimensional data and they won’t give the “same” answer. Also, there’s a subjective component on top: you are determining by eye where to cut an agglomerative tree and how many groups are in the PCA.

0
Entering edit mode

Actually my aim was to see if after batch effect removal the samples clustered as desired according to the two sample types. As can be seen from the hclust results and from PCA 63% variance is explained by PC1. So I guess the job has been correctly done.