Yes, I'm pretty sure that they chose the DMFs based on comparisons between LC and HC, and ME/CFS and HC. So, I think the PCA is primarily an illustration of their DMF selection process rather than showing any truth about the data.
Warren also seemed to think that the consistency of the results within the groups as shown by the close clustering of the individuals in each group on the PCA chart said something real about the data. But, the researchers specifically selected fragments where the P value for the comparison between a disease cohort and the controls was good. Intuitively, that seems likely to me to produce clustering within groups.
Even if I'm right about this, does it matter? I think it does. It's a bit like the subjective outcomes in unblinded trials recipe - the process, I think, is a guaranteed way of getting a certain outcome. But, it doesn't actually mean anything, it misleads and, it affects credibility. And it leads people to make faulty conclusions e.g. about how different the LC and ME/CFS groups are.
I'd love to get hold of the data for the 70,000 odd fragments, and then reallocate the members of the matched trios across three new random groups, select DMFs in the way the researchers did (but for the new groups) and see if fairly similar PCAs could be produced. I think that they probably could be.