Present introduction to project | Bill Swope | Presentation on work looking at conformer energy ordering issues in public dataset Lee-Ping said that decision was made not to include ddE in FB targets, but it could take it Current pipeline primarily looks at structures, not surprising that we have some issues with energies
|
Present progress so far | Pavan Behara | Slides Weight targets based on QM and MM energy for Torsions Use energy for optimized conformer geometry too ddE target--good energies, but RMSD from QM high RMSD as objective--good structures, bad energy order Combine the two--gives ok energy order and structure Restrained minimization works pretty well for both ddE and RMSD Energy/force matching for opt geo targets Relative energy of all pairs would be next step Iterative fitting works in some cases--opt FF, generate MM minima, generate QM, add to training, opt FF again, repeat--works in some cases. Wouldn’t help if issue was due to non-bonded terms. May want to record outliers and check typing
Then repeat Bill’s analysis on new FF
|
Discussion of status/how to proceed |  | BS: Should think about what is the appropriate distance metric--is RMSD > 0.4 A too high? Should it be size-dependent BS: Have you tried changing anything except valence (bonds, angles?)? Have you looked into whether it could be LJ, Charge, etc? LM: Any issue with parameter typing? BS: Haven’t classified it that way but could be useful PB: Could help but better to focus on targets, easy to get stuck on focusing on outliers
BS: How many cases where different conformers have different electronic structure? Would we need to look at a different fixed charge model? Our charge model may be favoring certain types of conformers LW: Hard to decouple bonded/non-bonded parameters, and expensive to refit non-bonded BS: Using forces seems like it would help for matching DFT surface PB: Agrees. Also, we’re not fitting vibrational frequencies--may need to think about that LW: Why no vibrational frequencies? PB: Fit was good without it BS: How specific are frequencies to conformers? Primarily care about bond stretch, angle bend, high frequency that isn’t sensitive to conformers
LW: How is working with Chapin to look at pairwise small molecule energies? Seems relevant Different objective function for torsion optimizations PB: Tried fitting just torsions, not better than Sage, now trying to fit bonds/angles too
Where to start? PB has sample branch of FB with relative energy targets Best to start with new target options To change targets, need to modify FB code, make new target within FB
BS: Problem with workflow is that it doesn’t take into account atom symmetry, e.g. penalizes rotating a methyl group to an equivalent position Doesn’t affect RMSD too much, but would be good to fix it RDKit has RMSD function that takes into account symmetry PB just using BS code so hasn’t fixed that
|