Conducting per-feature analysis


Wanted to raise a discussion about pros and cons for different methods for testing a feature association with metadata column (e.g. change of any ASV's in my samples with change in weight during a study).

The way I see it, there are 3 main options:

  • Using linear mixed model, for example from QIIME longitudinal plugin. From my understanding there are 2 major limitations: associations between microbial features and metadata columns would probably not be normally distributed; and I don't if there is an option to "run" over each feature in a feature table to check the model for each feature separately (and correct for multiple comparisons obviously).
  • Using a popular regression package for microbiome data, for example Maaslin2. Major downside is that it is not in QIIME :frowning:
  • Using one of the machine learning prediction models QIIME offers, which is a bit different ballgame and doesn't provide P (or q) values.

Would love to hear people opinions and experiences of why choosing one method over the other, some other pros and cons, and other methods for testing feature-outcome associations.


1 Like

Having the same tasks and considerations I ended up using Maaslin2 package outside of Qiime2. It is relatively easy to export all necessary files from Qiime2 to the Maaslin2 compatible format.

1 Like