I moved this to a separate discussion as it was getting away from the original topic.
I’m afraid I don’t have a straight answer for you, nor do I really think there is a straight answer here. I’m sure everyone has their own approaches, and those will be largely driven by the type of data they have, and not seldom, based on convenience.
The choice of DA tools is complex enough when talking about microbe read counts, but the added layer of producing a new, much bigger, prediction table from those counts is even less explored. But assuming that all the DA tools are fit for predicted tables as well, here are some general thoughts.
There are some key differences amongst the various differential abundance tools out there, of which there are plenty, so sometimes you can choose the right tool based on your needs. For example, ANCOM makes some assumptions about the # of features being different between your groups (if I recall correctly that’s <25%) and it tends to be one of the more conservative tools there are, this is in contrast to DEseq2 which in my little experience tends to be too liberal. Whether it is appropriate or not for compositional microbiome data is a whole other discussion. But I don’t think you will get any back lash from reviewers from using either of these tools. I’m not sure about the assumptions for newer ANCOM varieties. Other tools like Aldex2 and songbird have their own strength and quirks, with songbird requiring a lot more user input and actually not doing any statistical testing on its own, but rather it will help you choose important log ratios with ranks. The q2 score being too low is not ideal but it is not an absolute test that tells you whether or not you should be able to move forward or not. As long as its not negative I think you can proceed and just be cautious in your results interpretation. If you are able to optimize that score by playing around with the hyperparameters, that’s great, but don’t let it stop you from exploring your data further. I often like to try more than one tool at a time and see if they agree. If they do, then that’s reassuring, and if they don’t then those features were on the fringe anyways and you likely lack the power/confident in your tests. As for the memory issues you are experiencing I’m really not sure but that sounds really excessive, I’ve never had issues like that with ALDEx2. I bet you can significantly reduce the memory usage by doing some decent prevalence/frequency filtering which are recommended anyways for these types of analysis. If you continue to experience issues then I’d recommend making a new post with details about your table and we can get the developers’ input on that.
Also, here is a recent preprint on this topic that may be helpful. Sorry I couldn’t offer any direct help. Perhaps some others would like to qiime in.