PCA interpretation

Question

[Irving]

I performed a PCA with sequencing data. There are two treatments and two replicates for each treatment. I'm confused about how the samples are grouped, and especially about the percentages on the axes of the plot — what do the 98% and 1% mean?

[jared.andrews07]

How did you do this PCA and make these plots? The axes labels for variation make no sense (and thus I suspect they were swapped).

[Irving]

I used the following script:

BAM files

R1="DMSO_H1_noduplicates_sorted.bam" R2="DMSO_H2_noduplicates_sorted.bam" R4="TPL_H1_noduplicates_sorted.bam" R5="TPL_H2_noduplicates_sorted.bam"

Output file

OUTPUT_DMSOTPL="multiBAMsummary_DMSO+TPL_k562.npz"

Compute read coverage bins across the genome

multiBamSummary bins -b $R1 $R2 $R4 $R5 -o $OUTPUT_DMSOTPL

Perform PCA and generate plot

plotPCA -in $OUTPUT_DMSOTPL -o plotPCA-BAM_DMSO+TPL.png

[swbarnes2]

You are trying to PCA these by what, exactly? Binned coverage? Is this really informative?

[rfran010]

Assuming this is deepTools, I recommend reviewing the documentation and choosing one of the two indicated options:

Named Arguments
--transpose
Perform the PCA on the transposed matrix, (i.e., on the matrix where rows are samples and columns are bins/features. This then matches what is typically done in R.

--rowCenter
When specified, each row (bin, gene, etc.) in the matrix is centered at 0 before the PCA is computed. This is useful only if you have a strong bin/gene/etc. correlation and the resulting principal component has samples stacked vertically. This option is not applicable if –transpose is specified.

[ATpoint]

I do not recommend exploring noise-prone assays like ChIP-seq in a "blind" matter like this, using just the BAM files as input. Do proper peak calling, then make a consensus peal list, e.g. by merging all peaks, then a count matrix using featureCounts, then a PCA in R. Everything else, for me, is just superficial and not robust. Lots of threads here on merging peaks and all that.