I’m wondering and looking for feedback on how to interpret mmvec biplots. Is there a way to determine whether the associations shown are strong. Often there are so many metabolites associated with so many microbes its hard to know what to report. Just seeing if others have used these plots.
Hi @quinnr, the way to think about this figuring out how to partition variation in the metabolites with respect the microbes. If you look at co-occurrence plots, it is difficult to see which microbes have the strongest relationship with regards to the metabolite abundances. Biplots can help you cut through the noise and see the top microbial drivers of metabolite variation.
The cystic fibrosis example is the one that I like to use to showcase this. From the biplot, you can see there are really only two groups of microbes, which turn out to be largely characterized by anaerobes and pathogens. The longer the arrows are, the more variation those microbe explain.
There is another extensive discussion on how to interpret biplots that maybe worth checking out
Thanks Jamie, this is helpful. Another question related to this is do you have a way of ‘scoring’ an overall relationship of microbiome/metabolome with mmvec. Is there a good way to report overall model fit? i’m wondering if we find associations between microbes and molecules, but overall the signal is weak, that may have particular implications for our system. Thanks!
Hi @quinnr, this is a great question. We recently added in support for null models in the last mmvec release (see here: https://github.com/biocore/mmvec#null-models-and-qiime-2--mmvec).
This will allow one to determine how well microbial reads can predict metabolite abundances in held-out samples, and compare this against a null model.