Diversity estimates on FeatureTable[RelativeFrequency]


I am trying to perform ordination based on beta-diversity estimates, ideally weighted UniFrac. I would prefer not to rarefy, but to use simple proportion normalization (relative abundance). I can convert my feature table of counts (FeatureTable[Frequency]) to a feature table of relative abundances (FeatureTable[RelativeFrequency]), but none of the beta diversity (or phylogenetic beta diversity) calculators in q2-diversity seem to accept FeatureTable[RelativeFrequency].

Is there a way around this? I can think of two options, both of which involve creating a relative frequency table, then lying about its type and pretending it’s a FeatureTable[Frequency]:

  1. Convert to FeatureTable[RelativeFrequency], then unzip the archive, edit metadata.yaml and change type: FeatureTable[RelativeFrequency] to type: FeatureTable[Frequency]
  2. Export my FeatureTable[Frequency], do the proportion normalization myself, and re-import to FeatureTable[Frequency].

To my surprise, option (1) seemed to work for Bray-Curtis using qiime diversity beta, but not for weighted or unweighted UniFrac using qiime diversity beta-phylogenic (the resulting distance matrix was all zeros)

Is there a better way to do this?

Hey there @casey.grun!

That is correct.

That won’t actually change the semantic type of the data, and will probably just corrupt your QZA.

If you wanted to go this route, the “right” way would be to export the FeatureTable[RelativeFrequency], then to reimport as FeatureTable[Frequency] (although, this will reset your provenance, which is a bummer).

The results should be the same for unweighted unifrac either way, right? Since that is a qualitative metric the only thing that matters is presence or absence of a feature, which means that normalized or not, the results will be the same, right? The same is true for any qualitative metrics — they aren’t sensitive to abundance, they are driven by presence/absence.

As far as Bray-Curtis goes, this paper seems to indicate that proportions work well with this metric, but not the others (credit: @Nicholas_Bokulich).


This topic was automatically closed 31 days after the last reply. New replies are no longer allowed.