OTU impacts on PCA

[BuildableDuck (Ian Tannenbaum)]

I am using Qiime2-2019.1 and don’t see any way to determine which OTUs are impacting the PCoA. Is there such a command in Qiime2?

Cheers

[jwdebelius (Justine Debelius)]

Hi @BuildableDuck,

That’s a complicated question! My short answer is no, there’s no series of easy commands to tell you the answer. This isn’t a failing of QIIME 2, it’s a challenge of the way we compress and visualize the data!
The longer answer is, “maybe”.

Unforuntately, PCoA isn’t like PCA where you can do a factor analysis and determine the individual features driving the separation. That information is lost in the distance transformation and cannot be retrieved. (It would make life so much easier if it could.)

If you’re working with a weighted metric (Bray Curtis, Weighted UniFrac distance, pick your favorite…) you could try a biplot in Emperor to see if a specific set of features localize. Im not sure about filtering parameters, etc, you might have to play around.

A second is to use a feature identification tool, like ANCOM, to identify features that differ between the groups and then seeing if filtering to these features re-capitulates your patterns via procrustes analysis.

Best,
Justine

[Nicholas_Bokulich]

Also, see this:

Robust Aitchison PCA Beta Diversity with DEICODE

DEICODE (pronounced like decode /de.ko.de/) Documentation available in the pugin library. DEICODE is a form of Aitchison Distance that is robust to high levels of sparsity. DEICODE utilizes a natural solution to the zero problem formulated in recommendation systems called matrix completion. A simple way to interpret the method is, as a robust compositional PCA (via SVD) where zero values do not influence the resulting ordination. One of the benefits of using DEICODE is the ability to reveal …

[BuildableDuck (Ian Tannenbaum)]

Thanks for the responses, that DEICODE package did exactly what I needed. Thanks for all the help.