Balances in Gneiss

Hi, I’m trying to perform Gneiss analysis on my data and I’m seeing a great number of balances in my output. There are several things I’m wondering about:
(1) Is there any guideline as to which balances to pick and look at?
(2) I tried to look the the p values. But in one of my analysis, I got quite a lot of balances with a significant p value. Is there a recommended way for me to piece them together?
(3) I’m looking at the effects of a particular treatment but four experiments are done on four different days and time is supposed to be another variable. Does the result I get from balance-taxonomy control for the effect of another variable? Or is the other variable simply collapsed?

For question (1), should I primarily be looking at the balances that account for the most variance (like y0 and y1)?

Also, in general, which method of composition analysis is better, ANCOM or Gneiss?

Why not try both?

Concerning the other questions

  1. Not really - we try to provide information in the dendrogram-heatmap, the balance-taxonomy in addition to the regression summaries to help figure out which balances could be meaningful. The only thing that I would caution against is picking balances that are close to the tips of the tree, since those values could be heavily impacted by low count noise.
  2. I’d recommend not trying to threshold p-values, but rather starting off with the most significant ones first.
  3. I don’t completely understand this question – maybe some output from balance-taxonomy could help clarify this question?
1 Like

Hi Jamie, thank you so much for replying.

Here's the regression summary I got.
regression_summary_metazoa_1.qzv (467.5 KB)
Here I'm mostly focusing on the effects of treatment (DMSO vs High HHQ, time zero data not included). Should I be looking at the balance that is significant but at the tip (y263) or the balances that are not quite significant but seem to account for the most variance (y0 and y1)? Also, when I'm looking at the taxa bar plot, there seems to be a difference between treatments but here Gneiss results seem to disagree with that?

I just tried with ANCOM and it tells me that I have no significantly different features. One thing I'm not sure is how to determine whether I should expect no more than 25% of the features to be different, the assumption of ANCOM.

I think a better way to phrase my question is that, is there a way in Gneiss to check for the interaction between treatment and experiment?
y1_taxa_summary_metazoa.qzv (129.1 KB)

Also, is it possible for me to look at just one experiment and only consider the effect of treatment? Is it possible to use Gneiss with only a discrete variable in that way?

Thank you so much!

Hi @ygao2, I’m still having a bit of trouble parsing your question.

What do you mean that the Gneiss results / taxa bar plots disagree?
The taxa bar plots are just summarizing the breakdown of the regression results.

Also what specific hypothesis are you trying to test for? Are you comparing treatment vs control for just one experiment?

The easy way about this is to just filter out samples belonging to other experiments and rerun. You can definitely handle discrete variables in Gneiss.

Looking at your results, it looks like Treatment[T.time zero] is significant everywhere. Not sure if this is the specific hypothesis you are testing for. See my blog post here on designing formulas using patsy – the same formula can be copied/pasted into the Gneiss --p-formula option.

And yes, you can definitely look at interaction effects, its just

--p-formula treatment * experiment

1 Like

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