AM1 Restraint Studies

Previous Connor Davel meeting notes: https://docs.google.com/document/d/1gSwKIUWmJkWRyrRNXm2lwH5ZIF752TiQOXwuyZhm728/edit?pli=1#

Relevant section/project abstract draft:

It’s necessary to preserve the connectivity of a molecule when performing AM1-Mulliken calculations to assign partial charges to its atoms. There have been several reported cases of very poor-quality charges coming from the Open Force Field Toolkit due to molecule rearrangement during AM1 geometry optimizations. In this project, we will study restraint schemes that can be applied on top of AM1 calculation that allow a molecule to relax to its minimum energy geometry, but prevent it from making or breaking existing chemical bonds.

Motivating issue:

OpenEye theory documentation for AM1BCC: https://docs.eyesopen.com/toolkits/python/quacpactk/molchargetheory.html#am1bcc-charges

Current roadmap:

	Problem defined	Proof of concept implementation	Prototype open-source implementation	Production-ready
Database of 10+ molecules that will experience proton transfers
Database of 10+ Non-zwitterion molecules that experience proton transfers
Heavy atom rearrangement dataset
Way to detect proton transfers
sqm implementation that doesn’t allow nuclei to move
sqm implementation with “light restraints”
AM1-GeomeTRIC interface
Testing AM1-GeomeTRIC restraint scheme implementations

Bill Swope 10:36
Jeff; after a good night's sleep and a clear head, I re-read your detailed posts from yesterday. Thanks a lot for taking the time to write them out. Very clear and detailed! I should redo my dipole calculations making sure I am using the charges that were actually in the benchmark optimizations, and get some statistics on how many zwitterions there were and what fraction of them messed up due to the proton migration.

Jeffrey Wagner 12:47
I'm glad it was useful and readable -- Communicating complex technical info is not my strength so this probably took some effort to read.I'd love to know how frequently these proton migrations happen, if your dataset reveals that. Or in general, how frequently AM1-mulliken gives a radically different result than b3lyp-d3bj-RESP. If you're going to dive deeply into that problem then Simon may also be interested in following!

Tuesday, May 11th

Bill Swope 16:22
Jeff, here is some report about what I am finding. (I think it might be good to link Simon into the discussion - good suggestion.) The charge migration is most likely to happen with zwitterions, so I extracted a subset of 40 compounds that (1) had two formal charges indicated in the sdf file (CHG lines), and (2) were overall neutral. These compounds ultimately produced about 300 conformers in our benchmark data set. Note that we could also have charge migration during the AM1 optimization in other cases, such as where there were three formal charges, (+1, +1, -1), but I omitted these because I was monitoring with dipole moment calculations. Comparing charge models produced with "from_mapped_smiles" with "from_smiles" I see that some results are identical, others are very close (but not identical), some are much better and some are much worse. The most significant indicator for me that something has gone wrong, since all I have access to is the resultant charges, is that if I put the model charges on the b3lyp-optimzed structures and compute the dipole moment, I get something WAY off from the b3lyp dipole moment. After the hydrogen migration, AM1 would be giving us charges that might be appropriate for uncharged (non-zwitterionic) form, but are poor for the zwitterion. In most cases, then, the dipole moment computed with these charges result in severely depolarized systems, with dipole moments too small by as much as 12 Debye.