What is the recommended method for specifying enriched or depleted features?

[jolespin]

I'm following the tutorial here:
https://birdman.readthedocs.io/en/stable/default_model_example.html

With other software like ALDEx2 and edgeR you can specify a cutoff using FDR. I understand that working in a bayesian framework is different so I have a few questions:

• What is the recommended way for automating the feature sets using differentials?
• Is there an analog to an FDR cutoff that could be used that is generalizable to every run?
• Does this require manual curation on a case-by-case basis to determine which differentials are statistically enriched or depleted?

[gibsramen]

Hi, @jolespin

Great questions.

Overall, the goal of the approach is to eschew the traditional FDR/p-value cutoffs for determining features of interest. We tend to follow the work of Morton 2019 & Fedarko 2020 making use of ranks. In the Bayesian context, I tend to use the posterior means of the features for ranking purposes.

In terms of something analogous to an FDR cutoff, I usually take the top and bottom 10% of features as ranked by posterior mean for use in log-ratio analysis but this is an arbitrary threshold. Determining reference frames is still an open question. One could automate the process of determining the top and bottom X% of microbes (and I actually do something similar in qadabra). An alternative is to use a priori biological knowledge to determine a reference frame.

Feel free to follow up via e-mail as I am more responsive than here.