Hi @seibikobara - there are multiple avenues to use MelonnPan and not surprisingly, some of them are easier than others.
First of all, if you have paired metabolites and microbiome data, you can always use the MelonnPan-Train workflow and apply the trained model to new microbiome samples to predict metabolites. This particular avenue does not depend on Uniref90 annotation and likewise, is reference-agnostic.
Second, for the situation above when you have paired metabolites and microbiome data, you can also perform function prediction using PICRUSt and then use the PICRUSt output as input to perform metabolite prediction using other tools such as MIMOSA or using MelonnPan similar to the first avenue mentioned above.
Third, if you don’t have paired metabolites and microbiome data, you may perform function prediction using PICRUSt but it is not possible to convert KOs to UniRef90s since that mapping is one-to-many as discussed here.
We will have a future MelonnPan version that will eventually support non-Uniref90 input, but as of yet, MelonnPan is limited to UniRef90’s for prediction using the default trained model or without re-training a model.
All the best,
Himel