I’ve written a little bit about this here:
The short version is that we have not done a lot of testing of our methods on long reads yet, but they should be applicable to them in theory. The main constraint will be adjusting alignment to behave in a more “local” way (with short reads you expect a whole read to align inside a gene, but this may not be true for long reads, such that local alignment will be more tolerant of only part of the reads aligning).
More specifically, this means telling MetaPhlAn and HUMAnN to run Bowtie 2 in --very-sensitive-local mode (rather than the default --very-sensitive mode) and relaxing the query coverage settings in HUMAnN from 90% to something smaller (depending on the size of your reads and how long you’d like their alignments to be, e.g. if your long reads are ~1 knt and you’re OK with alignments >=300 nt then you’d set these thresholds to 30%).
Good luck and please let us know how it goes! 