Is ANCOM-BC2 the best choice for my data?

User Support
, , ,
1

My study is comparing the fungal diversity between soil that experienced a wildfire (NoRxF) between soil that was treated with prescribed fire prior to experiencing the wildfire (RxF). It's a paired study design where 6 sites from each "treatment" are paired for a total of 12 sites. I also sampled at two different time-points, 6-months and 12-months post-wildfire, as well as took 9 soil samples from each site using a 3x3m grid for a total of 216 samples (108 per time point). Here is a picture that should help communicate the study design:

PosterMap
PosterMap1920×3162 1.27 MB

My hypothesis is that the soil treated with prescribed fire will have a higher fungal diversity than the soil that only experienced the wildfire. To test that hypothesis, I'm looking to compare each of the paired sites. I also want to compare the sites between time points to see how the communities, and their differences, have changed.

I've read the ANCOM-BC2 paper that shows that ANCOM-BC2 is appropriate for categorical treatments, and I understand that Qiime2 uses the conservative sensitivity score (SS) version. I see that folks use the formula parameters to do paired comparisons, but I have a column called "Sample-Time" (where samples from site 5 collected in January are '5-1' while samples collected in July are '5-7') that I'm using to group the data:

## Calculate differential abundance
qiime composition ancombc2 \
  --i-table table_decont_pvlnc_unclst_noctl.qza \
  --m-metadata-file metadata_noctl.txt \
  --p-group 'Sample-Time' \
  --p-fixed-effects-formula 'Treatment' \
  --o-ancombc2-output ancom_pdecon_unclst.qza

As for the actual comparison between time points, a form moderator recommended that this user instead consider using Gemelli and construct an additive log ratio (ALR) using the factor loadings in order to account for the effect of each individual subject (for me, site).

One thing I should add is that my sampling depth varies a lot, and while the alpha rarefaction plotting showed a leveling out of Shannon diversity at 1556 and observed features at 6222, sampling at a depth of 6222 means I retain only 4.03% observations in 75.80% samples. I believe this means I'm losing A LOT of data, which makes metrics that don't need rarefaction (like differential abundance) more attractive.

Here is my data after filtering out sequences with <500 reads, extracting the ITS1 region with ITSxpress, denoinsing with Dada2, assigning taxonomy using Colin's 99% pretrained UNITE classifier, and decontaminating with Decontam:
asvs_decont_pvlnc_unclst_sgl.qzv (1.2 MB)

3

Is q2-longitudinal appropriate for only two time-points? I believe that's what user kbenmd meant by "before/after" in their post that this reply was split from. Unfortunately, that wasn't acknowledged yet in the thread. I tried it, and the plots looked gross:

Screenshot 2026-08-20 215749
Screenshot 2026-08-20 215749807×661 33.9 KB

UPDATE:
I've decided that the longitudinal approach is not appropriate for two time-points. Instead, I will be treating the resamples as a separate category instead of a numerical date. Now, I will compare diversity between paired sites to see if treatment made a difference. Then, I will compare between "paired" time-points to see if the change differed between treatments.

4

Hi @kranderson ,
It looks like you are using the LME action in q2-longitudinal which technically yes could work for 2 timepoints but really shines when you have more data points and are actually attempting a linear regression. Anyway, the visualization that you are showing is meant as a diagnostic anyway, not really a publication-ready plot, you should try out the volatility visualization or using q2-vizard if you want to create a nicer plot (though these also make most sense when you have more time points).

For two timepoints, I would suggest thepairwise-distances and pairwise-differences actions that were suggested in that other post, these are explicitly designed for the before/after design. But yes, treating it as categorical data is also an option (though take note of any assumptions that your statistical tests may have to decide if these still apply to your data)

Good luck!