I am using DEICODE on my 16S samples and I believe something weird is happening, as you can see in this plot - aitch-biplot-rarefied.qzv (1.7 MB). It seems that the vectors are somehow influencing the first axis of the PCOA and the variation I used to see is shifted to the second and third axes.
I re-did the PCOA biplot following the instructions in this other qiime2 post, and this was my resulting biplot- emperor-vis-biplot-pcoa-aitch.qzv (1.7 MB). For this last plot, I used aitchison as the metric for the qiime diversity beta step (to produce the distance matrix).
Why are the vectors of the DEICODE biplot interfering with the results of the ordination?
What exactly is the aitchison metric doing when I select it through the qiime diversity beta method? I don't really know what where to look for that info- as well as for the rest of the metrics here.
Are the plots generated from the two approaches supposed to be similar? Or should I expect some differences on the resulting ordinations?
The Aitchison metric is computed using pseudocounts (since you can’t take logs of zeros). This is expected to give you an arbitrarily bad bias. DEICODE has a more robust way to deal with missing values, so we recommend to use that one instead.
Regarding the first question, what I mean is that all the vectors are going in the same direction, and as a result, my samples are clumped on the left side of the plot. This is not seen on the plot that I produced with the other method, or any other ordination that I have produced before. I just want to know if perhaps this plot is supposed to look that way, or if my data have problems.
As for the second question, thanks for clarifying that. I know what decoide is supposed to be doing, but I didn't have enough information to understand what the aitchison metric implemented in the qiime diversity beta method is supposed to be doing. Again, if anyone knows a source where the metrics implemented in qiime diversity beta are explained, then we could do a more informed selection of the metrics. Anyhow, do you know if the aitchison metric does a ilr or a clr transformation? I just want to compare both methods, although I am more inclined to use deicode.
That I'm not sure about - that probably has to do properly centering the data. There may just be bias issues. @yoshiki may have more insights on this.
It doesn't matter -- the Euclidean distance between ilr / clr are equivalent (which is also known as the Aitchison distance). More can be found in Chapter 3 and 4 in the Modelling and Analysis of Compositional Data textbook (link here).
You can find a brief summary of sorts here. To read a more detailed explanation of each you'll have to research those on your own however as those concepts go deep into ecology and are a bit beyond the scope of a bioinformatics-focused environment.
If you’re in need a more in depth look at various ecological metrics, I would recommend the book Measuring Biological Diversity (Magurran 2004). It’s not perfect, but it does discuss most of the options currently available in QIIME 2 and their pros and cons on different types of data. I’ve found it incredibly helpful.
I think this may have been a centering issue with the biplot that was recently fixed in DEICODE v0.2.3. Could you try again after updating to the newest version? I think your problem should be fixed.
