Hey! This is such a great tool to use with QIIME2, thanks a lot. I wondered, if the sample metadata can also be used for the arrows to indicate what is driving the dissimilarity? If you know other ways to visualise this in PC(O)As, please share them, as other model output does indicate, but not visualise is (e.g. Adonis, beta correlation and group significance).
If I understand correctly, you are looking for a triplot PCA where you have samples as dots, vectors representing loadings (taxa), and other vectors representing your environmental variables (from your sample metadata). This is what you can get with a typical redundancy analysis (RDA) using R's vegan package, which is currently not supported in QIIME 2. But in order to get the triplot in R you need to operate on the raw count tables. If you were to use the rclr distance matrix as input, you would be running a distance-based RDA, which gives you samples by environmental factor biplots, but not feature (taxa) loading vectors. That's because you have already reduced your species count table into a distance matrix which doesn't retain the individual features.
There may be some way to implement rclr into R where you retain the original transformed table (not distances) and use that as input into RDA, but that's a whole other topic.
I will say however that I've had good results with running db-RDA using rclr and certainly worth a try.
Just a quick follow up, @jwdebelius reminded me that in the q2-gemelli plugin, which is built by the same developer who made q2-deicode, there is a
rclr-transformation action that you can use to essentially just rclr transform your regular feature-table then use that as input to your typical RDA which will then give you your triplots that would include all 3 aspects (subject, species, environment variables).