Discussion of charge methods for NAGL 2 | @Alexandra McIsaac | Recording Video Conferencing, Web Conferencing, Webinars, Screen Sharing Passcode: 8r5?Lxz= Notes were also shared with external folks on Google Drive here [I jotted the discussion points down during the meeting and fleshed out the discussion after--so it may not be in exactly the same order as the recording] JR: How is ESP calculated in RESP? Are there grid points inside the molecule? What is the grid like? JR: Typically ESP is looking around 2-2.5 VDW radius–does that capture H bonding radius? AG: Regarding ELF10/conformer averaging for RESP, could in theory do Boltzmann weighting to favor low-E conformers AG: Fitting RESP to dipole/quadrupole could alleviate buried atom problem LM: We could try that, or add those as fitting targets for GNN but keep regular RESP JR: GAMESS can fit RESP to dipole and quadrupole
AG: details of RESP vs MBIS ESP performance comparison? May not be looking at the same thing. RESP people have put a lot of thought into the grid. BS: Could MBIS be used to also parameterize the virtual sites? DM: Had a lab member who got HFEs similar to AM1BCC but without fitting anything, using MBIS BS: IPolQ is still ESP-based, likely would have same buried atom problem AG: Would fitting dipole at the same time help? BS: Could I guess, havne't looked. Still think it would be an issue
DM: Maybe solvent model is even more important than the charge model? AG: Could we use dimer/interaction energies for validation? ANI paper A neural network potential with rigorous treatment of long-range dispersion AG: Is conformer dependence good or bad? If the ESP is conformer dependent, why should the charges not be? LM: The problem with RESP is when the charges shouldn’t be conformer dependent, e.g. trans vs cis butene. The ESP will change in space, so the charges may not be the same, even though they often should be. MBIS doesn’t have this problem, but if there is a meaningful difference in the conformer that would produce a charge difference, that would be reflected. AG: There are situations where the conformer should affect the charge, and averaging over conformers or having fully conformer independent charges would neglect that.
DM: If we don’t want IPolQ, what’s next best option? BS: Could do nearest neighbor solvation shell around solute, could do QM optimization and get ESP from that DM: Wouldn’t next nearest neighbor affect structure too much? AG: What about implicit outside of first nearest neighbor DM: We’ve tried to do that but it depended too much on specific number of explicit waters, where they went, etc, depends on everything AG: maybe we don’t need full MD for IPolQ? BS: Important thing for IPolQ is averaging over many solvent conformations. Maybe 100s of calcs DM: Flexible solute? JR: No, solute fixed JR: What about one water for molecules with H bonds? LM: Would be concerned that that could have unpredictable effects
JR: are these the same charges as used for repulsive wall of LJ? DM: electrostatics are the only part we can derive from QM, so do that first, then re-fit the rest (LJ, valence) JR: Would need to think about 1-4 and 1-5 interactions
BS: Jorgenson uses CM1A, AM1BCC but for dipole/quadrupole, then scale up AG: How do we define success? LM: Typically assess solvation free energies (after LJ refit) and then if those are good move to ligand binding benchmarks DM: First metric of success is how well final charge/ESP reproduce ground truth QM (first paper), then build FF and do LJ refit etc
|
