Hi,
I tried ANCOM-BC to do some analysis at feature level. In some cases, I found the value of log-fold change was discordant to feature table. For example, I had a feature with the log-fold change value of -0.159 not favoring the experimental group. But when I inspected into the feature table, the experimental group actually had higher counts given that the control group had only zero counts in that feature. I wondered how I can explain about this result?
Thanks a lot!
Hi @b87401116,
That is weird! Would you mind sharing your table and ancom-bc results with us so that I can investigate?
@cherman2,
Thank you for the reply.
I found a mistake in my original post.
.... But when I inspected into the feature table, the experimental group actually had higher counts given that the control group had only zero counts in that feature.
should be corrected to
.... But when I inspected into the feature table, the experimental group actually had higher median counts given that the control group had only zero median counts in that feature.
In this feature, I think it is probably reasonable because the mean relative abundance was similar although the mean relative abundance was still higher in the exp group (0.00132) than in the control group (0.00127)
However, in the data provided below, the feature f7858c47d360f63a1dd7fb65c3489973 has log fold change of -0.160 which means the abundance should be lower in the exp group. The median abundance is 0.036 and 0.012 in exp and control groups respectively. The mean abundance is 0.045 and 0.012 in exp and control groups which is almost four-fold high.
I hope this puzzlesome problem will be solved. Thanks a lot
table.qza (63.2 KB)
metadata.tsv (140 Bytes)
ancombc.qza (73.5 KB)
Hi @b87401116
Sorry for the late reply! We are looking into this and we will get back to you ASAP!
Hi @b87401116,
Sorry for the late response on this.
So this question kinda boils down to the basics on compositional microbiomes and differential abundance. Also, thanks for asking this question, I learned alot trying to understand why this was occurring!
Microbiomes are more complicated than just one microbe increasing. An increase in one microbe doesn't mean that the microbiome has more microbes in it, it usually is a general shift in composition. I.e. If one microbe in a microbiome is depleted, other microbes are basically enriched in order to make up the compositional difference.
This is why when we are investigating differential abundance, we have to consider how other microbes in the microbiome are changing. This is why we can not just look at relative abundance of one microbe in the microbiome. In that one comparison, it may look like that microbe abundance is changing but we need to also consider how the other microbes in the microbiome are changing.
For your data, when looking at the relative abundance of f7858c47d360f63a1dd7fb65c3489973, you can see that the relative abundance does look like it is increasing in your treatment group, but ancom-bc's log-fold-change says that it is minorly negative. Basically this is saying that in comparison to other microbes and their shifts, f7858c47d360f63a1dd7fb65c3489973 is negatively changing. Other significant microbes are changing more, making the change in distribution of f7858c47d360f63a1dd7fb65c3489973 minorly negative (but probably more accurately no change in abundance)
TLDR: Microbiomes are compositional so comparing the relative abundance of a single microbe is not representative. Comparing a ratio like (abundance of microbe A/abundance of Microbe B) would be more accurate. However, this is why we have complex tools like ANCOM-BC to help us identify differential abundant microbes!
If you want to look into this more:
Figure 2 of this paper has a GREAT diagram of this:
@cherman2,
Thank you for the reply. It is somehow reasonable to have different results since the calculation of ANCOM-BC is not based on relative abundance. However, I still feel that the difference between ANCOM-BC and traditional methods is a little bet big. I will try to get some calculations on the ratios later. At least, I am now more confident that I did not make errors in the calculation step.
Hi!
In one of my datasets, I performed both LEfSe and ANCOM-BC. While most of the results were similar, one feature, reported by LEfSe as increased, decreased according to the ANCOM-BC. It looks like normalization matters indeed and there are disagreements between packages.
This topic was automatically closed 31 days after the last reply. New replies are no longer allowed.