I had a question regarding analysis of differential abundance of OTUs using DESeq2. As an example, I have 3 samples with 3 replicates each, and want to compare whether selected OTUs are significantly different across all habitats (similar to ANOVA). The count table below shows the setup. With the DESeq function (design=~habitat; where sample 1 all replicates= habitat1, sample 2 all replicates=habitat2, sample3 all replicates=habitat3) I am able to perform pairwise comparisons across the habitats (e.g. habitat1 vs. habitat2, habitat1 vs. habitat3, etc.), and with the LRT parameter (reduced=~1) can perform a global comparison across all habitats. But I assume this LRT treats every replicate separately and performs the test across all replicates of all samples in the three habitats (sample1_rep1 vs. sample1_rep2 vs. sample1_rep3 vs. sample2_rep1...vs. sample3rep3). But is there a way to group the replicates in some way first before running the LRT so that the test is not by replicate but is by habitat with grouped replicates? These are biological replicates so would prefer not to collapse by summing the counts as is recommended for technical replicates.

                             OTU1      OTU2       OTU3 

Sample1_rep1      5200       32508      1890                                                     

Sample1_rep2      356        52541        0

Sample1_rep3     2453       28167        3814

Sample2_rep1     11897     31699       0

Sample2_rep2     4690      49127        0

Sample2_rep3     4950      47731        0

Sample3_rep1     3925      6182         513

Sample3_rep2     3148     9783        362

Sample3_rep3     1241     6166         0

Thanks much,

Jesse Port

"But I assume this LRT treats every replicate separately and performs the test across all replicates of all samples in the three habitats (sample1_rep1 vs. sample1_rep2 vs. sample1_rep3 vs. sample2_rep1...vs. sample3rep3). But is there a way to group the replicates in some way first before running the LRT so that the test is not by replicate but is by habitat with grouped replicates?"

hi Jesse,

I think you're just off on the assumption of what's happening. A likelihood ratio test with full design ~ habitat and reduced design ~ 1 does group the samples by habitat. The meaning of the test is: is the increase in the likelihood of the model which groups the samples by habitat significantly more than the likelihood from the model which only fits an intercept term. The wikipedia page on LRT might offer some useful background: http://en.wikipedia.org/wiki/Likelihood-ratio_test