Number of features in the original database leads to different number of significantly discriminative features in LEfSe?

Hi all,

I am using LEfSe web version.

My whole database includes 899 features (in rows) and 8 observations (in columns, 4 observations per group). I used LEfSe to identify significantly discriminating features between the groups but I used the data in 2 different ways:

Way 1: I used the whole dataset, i.e. I included 899 features for the 8 observations, and the result was this:

Way 2: I used a subset of my data, i.e. I included only 13 features for the same 8 observations, and the result was this:

I have 2 questions:

Q1: Why is the LDA score for Actinomyces around 3 in way 1 but it’s 4.8 in way 2?
Q2: Why did I find 6 significantly discriminative features in way 1 but 8 in way 2?

While creating this post, I also noticed that step B results differ, if this helps:

Way 1:

Way 2:

Since I don’t have subclass, I am also wondering about the results from step B way 1, why do I have a different number of significantly discriminative features before and after the wilcoxon test? I am asking this because I have seen solutions for people using LEfSe in python or R, but I didn’t see anything about the web version. I am supposing that in the absence of subclass, the wilcoxon text is made between the observations, am I correct?

Many thanks,

Hi Joana,

For your first question, I believe you are getting a larger LDA score because with the reduced features the LDA is able to be more discriminative between the two groups, but I could be wrong about this and it is hard to guess 100% w/o looking at the raw data.

For your second section. To me it looks like the Wilcoxon test isn’t filtering out an additional features (as it shouldn’t when you are running without subclasses), which is why it says 10 (10). You are losing features because you have fewer features in the first run with a LDA score of greater than 2, they are significant but they did reach the LDA threshold. That is a moveable feature on-line if you want to change that filter threshold.