ANCOM result not matching up with alpha and beta diversity results

Hello Qiime2 geniuses,

I am baffled by the fact that I got significantly reduced species richness (Shannon’s index and Faith Phylogenetic Diversity) and evenness and also significantly different PERMANOVA UniFrac distance in my treatment compared to my control but when I did the ANCOM, only one taxon (a bacterial family that is minor constituent of the samples) significantly decreased in the treatment relative to the control. How can that be? Should I assume that the ANCOM results are invalid?


Hi @Lisa_Crummett,

I won’t necessarily jump the gun saying that the ANCOM results are invalid. But it does require a little bit of investigation - it is possible that some of the assumptions about few changing species were violated (ANCOM requires at most 50% of the species are changing).

Do you have specific qzvs for us to help diagnose? May also be worthwhile to contact Siddhartha or Shyamal (the first and last authors on the ANCOM paper. They’ll likely have a better idea about the underlying statistical assumptions.


Thanks for your quick reply Jamie. I have attached the .qzv files for the ANCOM results and the alpha and beta diversity results for you to see. Be sure to have "treatment" selected in the drop down menu for the alpha and beta diversity results. All results are associated with days14-20 of my experiment and pH 7.6 is the treatment and pH 8.0 is the control. Let me know what you think. I will also try reaching out to the authors of the ANCOM paper. I am just in a mad rush to get a paper submitted with these results and I'm not sure how long it will take them to respond. Fingers crossed...

Thanks again,

Day14to20-Level6-ancom-Treatment.qzv (45.0 KB)

faith-pd-group-significance_days14to20.qzv (278.4 KB)
shannon-group-significance_days14to20.qzv (278.2 KB)
evenness-group-significance_days14to20.qzv (278.1 KB)
weighted-unifrac-Day14to20-Treatment-significance.qzv (243.8 KB)

Alpha diversity, beta diversity, and differential abundance tests are all looking at very different things. So there is not necessarily any contradiction in your results (though it is certainly worth examining very closely to make sure!).

For example: lower alpha diversity suggests the treatment is exerting some type of pressure on the community. Either some species are dominating the environment, or a range of other species are being “flushed out”, leading to lower richness and evenness. These are not necessarily the same species increasing/decreasing in each sample/group in response to the treatment!

I would expect a significant difference with weighted unifrac permanova to translate to at least some more significantly different features with ancom, but these tests are not directly connected so this is not a rule. However, there is a warning in your alpha and beta diversity visualizations that may explain this disparity:

Warning: Some samples were filtered from the input distance matrix because they were missing metadata values.
The input distance matrix contained 78 samples but permanova was computed on only 24 samples.

So ANCOM, alpha diversity, and permanova might not be tested on all the same samples! (it seems like ancom should have the same constraint but there is no warning message — I am not sure if it is silently dropping these samples or what, but maybe @mortonjt can elaborate). In the ANCOM visualization, there are different “group” names, which makes me wonder whether the same metadata column was used.

Let’s see what @mortonjt has to say!

1 Like

Thank you Nicholas for your feedback.
I understand what you are saying about the alpha and beta diversity results. There must be different taxa increasing/decreasing across samples within a treatment. Although I don't understand how you can get weighted UniFrac distances being greater between treatments than within treatments and not have more taxa having signif different relative frequencies across treatments in the ANCOM result. The warning that you see is correct because the metadata file that I used was filtered to remove days 2-12 as I wanted to focus on days 14-20. I double checked to make sure that the metadata file that was used had the correct "group" names and it does so that should not be a problem. I have attached the metadate file that I used and the table file.

Thank you,

metadata_Exper.Day14-20.txt (481 Bytes)

comp-Exper.Day14to20-table-taxonlevel-6.qza (47.9 KB)

Agree with @Nicholas_Bokulich, these methods give very different results.

Re the Unifrac distances - Unifrac is not a compositional distance, so
small changes of proportions won’t impact the distance too much (even if it is a large log-fold change). In fact, its possible to have only a handful of distinct species that change a lot in proportions, which will cause a large shift in the Unifrac distances.

From your ANCOM results, it looks like your few taxa are changing are expected to change a lot (W=160 is quite high if you only have 194 species). However, it may be a good idea to sanity check the thresholds with Shyamal – for instance, it would appear to me that k__Bacteria;p__Bacteroidetes;c__Flavobacteriia;o__Flavobacteriales;__;__ and k__Bacteria;p__Actinobacteria;c__Actinobacteria;o__Actinomycetales;f__Microbacteriaceae;g__Candidatus Aquiluna are also possible candidates, since their W scores are also quite high (and they really stick out in the volcano plot)…

Another good way to sanity check this is to pull out barcharts or heatmaps to sanity check those individual taxa.

1 Like

Thank you Jamie, that makes sense. I checked my taxa bar plots and confirmed that those two additional taxa that you mentioned (outliers in the volcano plot with high W values) are in fact higher in the control vs. the the treatment. I have contacted Shyamal to ask him about my results as well as the thresholds. Is there a Qiime2 command that would allow me to reduce the threshold so that I can view the quartile counts for these two additional taxa in addition to the W scores?



How did you determine that I have 194 species from the ANCOM results? I don’t see that value given anywhere?


I downloaded your ANCOM results and there were 194 rows in the table

And no, you can’t modify the threshold in qiime2, but you can tweak it in skbio - how to do this is something also worth asking Shyamal about

This topic was automatically closed 31 days after the last reply. New replies are no longer allowed.