Weighted_unifrac_normalized metric

[nerdynella]

Hi there!!

@Stephanieorch

--p-metric: weighted_normalized_unifrac

is the above --p-metric available yet? i am getting an 'invalid choice:' error when i pass it
Nsa

[Nicholas_Bokulich]

Use the command qiime diversity beta-phylogenetic --help (EDIT: or qiime diversity beta-phylogenetic-alt --help) to see all available metrics.

Looks like weighted_normalized_unifrac is available in beta-phylogenetic-alt. not an available metric (I think this list may be based on qiime1, not qiime2).

weighted UniFrac is normalized (by even sampling) in QIIME2.

I hope that helps!

[Stephanieorch (Stephanie Orchanian)]

Based on the “beta-phylogenetic-alt: Beta diversity (phylogenetic) - High Performance Computation” weighted_normalized_unifrac and weighted_unifrac are 2 different metrics. Should we remove the weighted_normalized_unifrac metric on both that page and the “Alpha and Beta Diversity Explanations and Commands” pages and make it clear that weighted_unifrac is normalized?

[gregcaporaso]

I think it makes sense to keep both of these metrics documented - there are differences in these methods, and both are accessible through qiime diversity beta-phylogenetic-alt.

[nerdynella]

@Nicholas_Bokulich i found it, thanks!

[wasade (Daniel McDonald)]

@Nicholas_Bokulich

weighted UniFrac is normalized (by even sampling) in QIIME2.

This is not totally accurate. Weighted normalized is normalized by the total branch length between samples in the phylogeny, whereas weighted unnormalized is not. The end result is that weighted normalized is bound [0, 1], while unnormalized is not. Does that make sense?

[Nicholas_Bokulich]

@wasade

Weighted normalized is normalized by the total branch length between samples in the phylogeny, whereas weighted unnormalized is not. The end result is that weighted normalized is bound [0, 1], while unnormalized is not.

Thanks @wasade — that makes sense.

So we are talking about two different levels of normalization here — by branch length, and by sampling depth (both methods are going to be evenly sampled in QIIME2 by default). Thanks for clarifying what "normalized" actually means in this context!

[wasade (Daniel McDonald)]

I just ran back through the publications and cannot find right now where this is differentiated unfortunately, but it was a distinction we made in QIIME1, and is one we make in Striped UniFrac. In Striped UniFrac, we actually get about a 50% reduction in compute and memory on for the unnormalized variant as the normalization is somewhat expensive, although that algorithm is fast enough that it doesn’t really matter…

On an aside, unweighted UniFrac is also normalized by the total branch length, but we do not expose an “unnormalized” variant. We could if there is interest or need.