`rppr reclass` should suggest updated classifications

[matsen]

This extends the work done in #171.

This will only be done with taxids that are less specific than a given rank. The default should be species, but this should be specifiable through a --max-rank flag.

For every rank that is is less specific than the max-rank flag, For every leaf that is not convex at that rank,

• say orig_taxid is the original taxid
• run guppy infer without the distance requirement, getting a new taxid new_taxid
• collect the following information into a CSV file:
  • seq_name: the leaf sequence name
  • old_taxid:
  • new_taxid:
  • old_name: the name corresponding to old_taxid
  • new_name: the name corresponding to new_taxid
  • makes_convex: does new_taxid sit inside the list of taxids that would make the tree convex
  • old_avg_dist: what is the average distance from the leaf to leaves maintaining their classification of old_taxid after the convexify step
  • new_avg_dist: what is the average distance from the leaf to leaves maintaining their classification of new_taxid after the convexify step
  • n_with_old: number of sequences with the old taxid
  • n_nonconvex: the number of the sequences with the old taxid that were called nonconvex by the convexify step

[habnabit]

This has been implemented in the branch for #192.

[matsen]

This is totally cool, and seems to work in the examples I checked out. Nice.

It would be wonderful to have a -t flag that would write out the discordance tree at the level convexified. This "suggestion" tree would indicate the new name. I think that a useful and easy-to-implement way to go would be to append the suggested name on to the sequence id. Thus if S001576771 gets reclassified to Pseudomonas monteilii then the new sequence id would be

S001576771 -> Pseudomonas monteilii

This would make it easy to review the reclassifications.

[matsen]

And one more thing. If we could have - replace any non-normal float that would be super. I.e. we shouldn't have any -nan in the avg distances.

Then an explanation of why we don't do the calculation in the docs... because comparing distances between ranks is not so meaningful.

[matsen]

This is going to be an ongoing project but this one and it's friend infer should go ahead and get merged into dev.