`edit accumulate` can produce tiny negative placement coordinates

[bcoltman]

Hi,

I’m seeing an issue where .jplace files produced by:

gappa edit accumulate --threshold 0.95

are rejected by guppy check / guppy fpd because some placement locations are very slightly negative, e.g.:

invalid placement location in <query_id>: -2.22045e-16 < 0
invalid placement location in <query_id>: -6.93889e-18 < 0

This occurs across all accumulated files in this dataset. The affected edges generally have normal branch lengths, for example:

edge 426  branch_length 0.257455
edge 406  branch_length 0.118157
edge 381  branch_length 0.881322

The values look like machine-precision deviations from zero rather than meaningful negative placement locations. My guess is that some accumulated placements are landing exactly at branch endpoints, and floating-point arithmetic/writing results in tiny negative distal/proximal coordinates.

Would it make sense for edit accumulate or the .jplace writer to clamp placement coordinates to the valid branch interval when deviations are within numerical precision?

I can provide a small failing .jplace file if helpful.

Environment/details:

gappa version: v0.9.0
command: gappa edit accumulate --threshold 0.95

[lczech]

Hi Benjamin @bcoltman,

thanks, indeed, that seems like a numerical issue. I've just pushed a fix to the dev branch - could you please check if that solves it?

If not: It might be that the issue is instead coming from the function that is writing out the jplace file again. In that case, please let me know if you still observe the issue, and if you could then also provide me with the small failing file to test?

Cheers and so long
Lucas

[bcoltman]

@lczech sorry for the late response. Thanks for implementing a potential fix so quickly. I'll test it and get back to you.

[bcoltman]

Hi @lczech, it did indeed fix the issue of negative placement locations. However, it then revealed an additional issue in downstream use which is that there can be tiny non-zero distal_length on zero-length branches, which are later rejected by guppy. I have opened that as an issue on Genesis lczech/genesis#14, as I believe that it is probably best handled there. However, I tested a fix locally and confirmed that the build resolved the issue. I have opened that as a draft PR #35, although it is probably best resolved upstream.