General updates | JW – Thanks for quick fix to CPU platform. I’m waiting for conda packages to come online so I can test again. JW – I worked a bit with Richard Gowers (head of OpenFE) to try writing a writing a more flexible PDB loader (one which can be taught about more residues). Results from hackathon: https://github.com/richardjgowers/dasPDB/. It seems like some big tradeoffs are Should we use atom/residue names exclusively to find chemical templates? Or can we use elements+bonds (detected by distance) to find templates? PE – Neither approach is really reliable. Using names is bad because lots of files not from RCSB use nonstandard names. And if you do anyhting other than standard residues then all bets are off. Using elements is hard because often the element column isn’t filled in, or contains info other than the actual element. Also bond detection by distance is unreliable.
JW – There’s also the MDAnalysis magic elements+connections → formal charge + bond order guesser PE – This will fail when explicit Hs may be missing Ions that could have multiple charges (technically possible to put in a PDB file, but very rarely used in practice)
JW – We also just excluded all radicals from OpenFF toolkit. This may make the MDAnalysis guesser feasible. PE – You’ll also need explicit elements and no missing atoms JW – PE, would you be interested in following/joining this work?
PE – When importing openff-toolkit, I get the advertisement for OE. I’d like that to go away. Would I cause any trouble by opening a PR to remove that? MT – We actually should have jut removed that in the most recent release. https://github.com/openforcefield/openff-toolkit/pull/1426 JW – PE, if you reinstall, could you verify this goes away? This should have been fixed in last week’s release (0.11.3) MT - tested in one of my envs and this is fine locally: (openff-interchange-env) [forcebalance] conda list | grep openeye | wc -l 13:08:38 ☁ fix-installation ☂
0
(openff-interchange-env) [forcebalance] python 13:08:44 ☁ fix-installation ☂
Python 3.9.13 | packaged by conda-forge | (main, May 27 2022, 17:00:52)
[Clang 13.0.1 ] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> from openff.toolkit import __version__
>>> __version__
'0.11.3'
>>>
MT – Picking a platform - A lot of our code explicitly requests Reference, what’s the automatic option? PE – It picks the fastest available, where speed is hard coded as Reference < CPU < OpenCL < CUDA. MT – Relative performance of Reference and CPU? PE – About ~10x, though for very small systems reference can actually be the fastest. Depends a lot between CPUs as well (n cores is a big contributor). MT – For us it’s mostly doing a large number of single points, with just a few simulations.
|