Dear group,
we have a MiSeq 16S-dataset featuring samples from enrichment studies, i.e. communities from a time series in which some OTUs become dominant over time, e.g. up to 90% of all reads. The biological question would to find a) which OTUs respond to different enrichment strategies and b) when they start to enrich. I guess, this qualifies as a expression analysis.
Thus, we need to normalize the data due to highly variable sequence depths (20,000 to 70, 000 reads) and to validate our post-hoc analysis.
I tried percentile-based normalization like CSS but i have just learned the hard way, that they are not suited for this dataset (as they typically want to see relatively invariate data). CSS, e.g., just took away all observations from the enriched OTUs until the enrichment effect was not visible anymore.
Rarefying is inadmissable as McMurdie & Holmes told us.
Total-Sum-Scaling (i.e. scaling to all reads in a sample) is dangerous because it is sensitive to compositional effects (as our samples tend to become very uneven over time).
Any ideas how to best treat the data would be greatly appreciated.
See link for some normalization options http://www.rna-seqblog.com/rpkm-fpkm-and-tpm-clearly-explained/
I am going to move this to a comment since the original question is about metagenomics and not RNAseq. I am not sure if this post is applicable.
Thank you, but our amplicons are all very similar in read lengths. It is the standard metagenomic 16S-survey.