`sourcetracker2 gibbs` generating empty output tables

[nick-youngblut]

I'm running sourcetracker2, version 2.0.1 on my data as with the following:

sourcetracker2 gibbs \
  --jobs 1 \
  -i counts.biom \
  -m samples.txt \
  -o tmp/ST_test/

... and the output mixing_proportions.txt and mixing_proportions_stds.txt tables only have a header. There is no error or warning. It would help to at least have a warning if the output table is empty. Any idea on what's going wrong?

[johnchase]

Hi @nick-youngblut is it possible to share your data, or a subset of the data that reproduces the problem? I agree that a warning may be appropriate here letting the user know what is going on, although I am unsure of what the the data looks like that is creating this issue.

[nick-youngblut]

After some checking, it appears that when I took a random subset of my entire samples table for creating a smaller test dataset, the random subset happened to not include a 'sink' samples. It appears that SourceTracker don't check for the presence of both 'source' and 'sink' samples in the samples table. Of course, this should be a basic requirement of the user when formatting the proper input, but if this happens, it is somewhat hard to debug, since SourceTracker still generates an output (just an empty table).

[nick-youngblut]

Just as an FYI, I've been assessing how removing rare features from the count table will affect the classifications. It appears that if a sample no long contains any features with a count > 0 (ie., sample total count = 0), then the resulting mixing proportions for that sample are 'nan'. A user may wonder why they are getting 'nan' values for all mixing proportions of a sample. Adding a warning when importing the count table into SourceTracker may help prevent confusion about this.