search

RosettaCommons / rosetta

The Rosetta Bio-macromolecule modeling package.
https://www.rosettacommons.org
Other
154 stars 63 forks source link

Test and fix for mp_f19_decoy_discrimination failure #49

Open roccomoretti opened 6 months ago

roccomoretti commented 6 months ago

The mp_f19_decoy_discrimination scientific test has a hard failure. It's crashing with the following issue in the minimization routine of relax. 

core.scoring.hbonds.hbonds_geom: [ ERROR ] NAN occurred in H-bonding calculations!
core.scoring.hbonds.hbonds_geom: [ ERROR ] Dxyz                    -NAN                    -NAN                    -NAN  Hxyz                    -NAN                    -NAN                    -NAN

This PR pulls out the offending run/structure as an integration test.

Playing around with it, I figured out two things

  1. The -constrain_relax_to_start_coords true flag is needed to show symptoms
  2. It was specifically the derivative calculations which were causing NaN issues.

Inserting a number of std::isnan() checks and trying to trace back the source indicates the problem was MembraneGeometry::g_radius_gradient_dz(), specifically where it did a (p.x()-xo)/std::abs(p.x()-xo)), which results in a NaN if p.x() equals xo. We can rearrange this slightly, not only to avoid the NaN, but also to reduce the number of calculations needed. While I was at it, additional efficiency changes in MembraneGeometry were also made.

Rituparnasamanta commented 5 months ago

The changes look good to me. elegant fix to the problem. Was the 90 test done once?

roccomoretti commented 4 months ago

Was the 90 test done once?

Yes. https://b3.graylab.jhu.edu/revision/rosetta:commits/20383

I noticed I didn't actually run the mp_f19_decoy_discrimination test on the server, so that's now queued as well: https://b3.graylab.jhu.edu/test/833171

roccomoretti commented 2 months ago

@Rituparnasamanta Any further thoughts, or additional corrections/tests I should perform?