Questions about the results of differential rank using Songbird

[Alex_Pham]

Hello,
First of all, thank you so much for developing qiime2 and other plugins, which are very helpful for my research. I would like to ask your help with the differential abundance analysis.
I used Songbird and Qurro for processing differential abundance analysis of 16S sequencing data of 5 environmental samples with 3 different treatments (15 samples in total).
It seems to be fine for the fit of my model. This is my command.
qiime songbird multinomial
--i-table filter-table-2nd.qza
--m-metadata-file sample-metadata.tsv
--p-formula "C(Source, Treatment('sludge'))"
--p-epochs 100000
--p-differential-prior 0.1
--p-summary-interval 0.1
--p-random-seed 3
--o-differentials Source-differentials.qza
--o-regression-stats Source-regression-stats.qza
--o-regression-biplot Source-regression-biplot.qza
--verbose
qiime songbird multinomial
--i-table filter-table-2nd.qza
--m-metadata-file sample-metadata.tsv
--p-formula "1"
--p-epochs 100000
--p-differential-prior 0.1
--p-summary-interval 0.1
--p-random-seed 3
--o-differentials filtered-null-diff.qza
--o-regression-stats filtered-null-stats.qza
--o-regression-biplot filtered-null-biplot.qza

qiime songbird summarize-paired
--i-regression-stats Source-regression-stats.qza
--i-baseline-stats filtered-null-stats.qza
--o-visualization Source-paired-summary.qzv

But visualization of the differential rank using Qurro shows only 5 OTUs were used to set up the model. Do you know the reasons for this result?
This is my command
qiime qurro differential-plot
--i-ranks Source-differentials.qza
--i-table table.qza
--m-sample-metadata-file sample-metadata.tsv
--verbose
--o-visualization Source-qurro_plot.qzv

Thank you for spending your time.
Best regards,
Pham

[fedarko (Marcus Fedarko)]

I think this has to do with your sample size -- Songbird's --p-min-feature-count parameter defaults to 10. This means that any features present in less than 10 samples will be removed (see this section of the README for details). Since it sounds like your study has 5 samples per group, this means that any features that are unique to just a single group will not be considered by Songbird (since they're only in 5 samples, not 10).

It may be possible to decrease --p-min-feature-count a bit (maybe to 8 or 9?) and see if this helps retain more features, but I'm hesitant to say that that is a good idea in general. This may just not be enough data to get much out of Songbird with, sorry

cc-ing @mortonjt in case he has any further thoughts on this.

[Alex_Pham]

Thank you for your information, Fedarko. I would have a better understanding now.
Do you have suggestions about the other differential abundance analysis that might work for small size samples?

[fedarko (Marcus Fedarko)]

I don't know of any tools that are explicitly designed to work well with small sample sizes... maybe ANCOM or ALDEx2 might be worth checking out? (Although I know ALDEx2 and Songbird are very similar under the hood, so the sample size may be a problem for that also.) This is beyond my expertise, sorry.

[Alex_Pham]

Thank you. I am trying these two methods.