Could q2-sample-classifier take relative abundance or CLR transformed abundance as input?

[05dongkeren (Nathan Li)]

Dear QIIME2 developer,

Thanks a lot for developing QIIME2 for us!

I would like to predict the sample metadata use the microbial data based on the random forest algorithm. Therefore, may I ask several questions about the q2-sample-classifier plugin?

1. Could I use the relative abundance (percentage) or the CLR transformed abundance (with plus and minus values) as the input?

2. In addition to 16S data, I also have the metagenomic data (genes, pathways) and metabolomic data (metabolites). Could I combine these datasets and use the combined file as the input for q2-sample-classifier? If so, do I need to perform any normalization or scale for all these datasets before running the q2-sample-classifier (because each dataset has specific data characteristics)? What do you suggest?

3. Could alpha-diversity indices (Shannon, observed features) also be used to predict the sample metadata through the q2-sample-classifier?

4. Could the shifts of certain taxon units (e.g. difference of an ASV between pre-treatment and after-treatment) be the input to predict the sample metadata through the q2-sample-classifier?

Many thanks for your time and guidance! I sincerely appreciate your help!

Warmest regards,
Nathan

[Nicholas_Bokulich]

Welcome @05dongkeren ,

@05dongkeren

Could I use the relative abundance (percentage) or the CLR transformed abundance (with plus and minus values) as the input?

No — right now the classify-samples pipeline only accepts frequency data as input. In theory other feature table types should be acceptable but this has yet to be tested or implemented.

@05dongkeren

In addition to 16S data, I also have the metagenomic data (genes, pathways) and metabolomic data (metabolites). Could I combine these datasets and use the combined file as the input for q2-sample-classifier?

Sure, you can use the q2-feature-classifier plugin to merge these tables and use as input to classify-samples. Just make sure that you have the same number of samples in each table, i.e., you do not have missing data types for some samples but not others.

@05dongkeren

If so, do I need to perform any normalization or scale for all these datasets before running the q2-sample-classifier (because each dataset has specific data characteristics)? What do you suggest?

Not for Random Forests. See discussion here:

Do I need to normalize (or scale) data for randomForest (R package)?

@05dongkeren

Could alpha-diversity indices (Shannon, observed features) also be used to predict the sample metadata through the q2-sample-classifier?

Sure! You can use the metatable action in q2-sample-classifier to merge alpha diversity (or even metadata) with your other feature data.

@05dongkeren

Could the shifts of certain taxon units (e.g. difference of an ASV between pre-treatment and after-treatment) be the input to predict the sample metadata through the q2-sample-classifier?

I am not sure how this would work out, or if the differences in relative abundances would be meaningful.

Good luck!