I am performing a study about differences in microbial communities between two sample types in an Antarctic ecosystem. According to PERMANOVA, there are significant differences between communities of the two sample types. However, when performing differential abundance analysis (Tried ANCOM and LEFSE so far) and performing Holm p-value correction, I detect zero significantly different features between sample types. I am wondering if it is possible that there is a significant global difference between two communities without any given ASV being the cause for that difference. Performing DA analysis at aggregated higher taxonomic levels gives similar results, with no taxa being differentially abundant.
I would be very thankful for any insight into this issue.
I think this could make sense. Your significant PERMANOVA result could be explained by small and consistent changes in the abundance of many different ASVs. None of these would be large enough to be statistically significant on their own.
Another possible explanation could be dispersion. PERMANOVA is sensitive both to intergroup variance and dispersion. In this regard, I found this post really useful: adonis, betadiver, and betadisp
If you don’t find DA ASVs, I wouldn’t expect to find DA species or genera. In my experience, the more you collapse by taxa, the fewer DA features you find (they tend to “cancel” each other when you aggregate).
I did run a PERMDISP analysis in QIIME2, and the results were not significant, which I understand means that I can fully trust PERMANOVA identifying the difference between two categories, correct? In that case, I can imagine that the variation for any given ASV is indeed too small to be detected by DA algorithms, especially considering I am testing upwards of 10000 ASVs at once for one dataset and 1000+ for another.
Another thing I was wondering if how to determine which DA algorithm is the best for a particular dataset. I went with ANCOM as a first choice (as it was the most widely used and recommended afaik), then switched over to LEFSE (as a sort of second opinion), and as I mentioned both analyses agree. However, I know that different algorithms can produce very different results (paper). I obviously wouldn't wanna try algorithms willy-nilly, but at the same time I can't help but wonder whether I am using the "most correct" method for my dataset.
That seems to be the case here! Anyway, I would also plot a PCoA just to double-check visually.
One small clarification: DA algorithms calculate a log fold change (or another measure of effect size) and the corresponding p-value for every single ASV you give them. So they are “detecting” all ASV variations. After that, you filter them using thresholds (e.g. logFC 2 and adjusted p-value 0.05) and decide that those are the significant ones. In your case, none of the values reach the thresholds.
Yes, multiple testing correction is also contributing to that.
I’m afraid I cannot help on that one. I usually do ANCOM-BC and I’ve never used LEfSe. What I normally do when I find several methods that produce the same output is to choose the one that methodologically looks better to me. In my case, it is ANCOM-BC, but everyone has their own opinion!
