I would like to know if there’s a way, having a distance matrix from, for example, a Jaccard metric, to get which features are the most “relevant” in the calculation of the distances for each sample,i.e. for each row or column of the matrix. Any suggestions?
The short answer is no, distance calculations are a reductive technique: you can’t tell which features truly drive the difference. There are some ways to get around it.
You’ve got a couple options. A bi-plot can help you look at the features that drive separation in PCoA space.
You could also look for things that are associated with the separation you’re looking at and see if you take a subset of the distance matrix how well that recapitulates your original patterns. (I use this sometimes when I want to look at a specific taxonomic clade.