2024-03-27 FF Fitting Meeting

Participants

  • @Willa Wang

  • @Brent Westbrook

  • Bill Swope

  • @Chapin Cavender

  • @Pavan Behara

  • @Trevor Gokey

  • @David Mobley

Goals

  •  

Discussion topics

Item

Presenter

Notes

Item

Presenter

Notes

optimization of a 3-point rigid polarizable water model

Willa Wang

  • slides / recording

  • BSwope: IMO the model to beat is OPC, not TIP3P

  • BS: is your model direct polarisations? AMOEBA is not

    • WW: iAMOEBA is direct, but has multiple expansions.

    • BS: some people only put polarisability on O. You’re putting it on all 3 atoms? It’d be interesting to see if you get improvement with just O being polarisable.

    • WW: both O and H have polarisability, so if I did, we’d probably need the charge to compensate for that more.

  • BS: water has a density maximum at 4 degrees, if you can get that with the water model, you have the correct balance of H-bond strengths between molecules. That’s a good target. TIP3P doesn’t have a density maximum – it just goes down with increasing temp.

  • CB: in the Summary slide, is the “This work” column optimization number 3?

    • WW: this column represents the best predicted property in any opt I’ve done – none have finished yet

    • CB: the charges seemed a bit high on some slides compared to TIP3P. You shouldn’t need these high atomic charges, since the polarisability should account for this at short range – I was assuming charges would be much lower. What’s your gas phase dipole?

    • WW: gas phase dipole is 1.8-ish. Calculated liquid dipole is around 2.2. That’s for charges optimized to ESPs [?], not to properties here.

  • CB: what about O distribution function? Does that have 3 or 2 peaks?

    • WW: not looked at this yet.

    • CB: worth monitoring as spacing of peaks should be tracking with density

  • CB: where are we mainly looking for differences in behaviour? I’m thinking structured waters around bound ligands in an active site, that we should expect polarisable water to have more correct behaviour

    • WW: could turn some of those scenarios into a target?

    • CB: could also use mixing targets, e.g. enthalpies of mixing of water with something nonpolar

    • LW – This is from the same dataset as the sage fit.

    • CB: in TIP4P-EW did you look at properties with nonpolar media?

    • BS: No, but Chris ? has done work on water models with interface properties, e.g. tendency for waters to orient near interface. RDF for TIP3P notably bad, could be worth looking at it

    •  

  • MT (chat) – It’s possible to fit to an RDF but not alongside other targets

    • WW – Do you have references where this was done?

    • MT (chat) – Not water models specifically. Will follow up later in slack - I don’t need to hold up this meeting

  • WW – Should I look at dimer/cluster eneriges?

    • BS – Dimer energies+geometries yes, clusters aren’t really needed. But the think you need to get rihgt are liquid properties.

    • CB – Agree. Could be good to use dimer energies+geos as an evaluation target before using them as fitting target.

  • BS – One reason why charges in non-pol model may appear to be higher is that your pol model will usualyl be getting polarized by env.

    • WW – My charges are lower than even iamoeba. It seems like hvap is extremely sensitive to my charges.

  • CBy – One reason for what you’re observing is that AMOBEBA and IAMOEBA are fully polarizable, and you’re just doing direct pol. So their polarization response is greater than yours. The non-pol models have to fold in the expected pol into their charges. So I’d think your direct model should expect lower charges than non pol, but higher charges than fully pol.

    • WW – I’m puzzled by high dielectric constant.

    • CBy – How does it compare to fully pol models?

    • WW – AMOEBA charges are lower in magnitude than mine.

    • CBy – So TIP3P has O charge of -0.8ish, yours has ~-0.7, and iAMOEBA is about ~-0.6. That makes sense given my explanation above. But I don’t understand why the dielectric constant is overestimated.

  • Cby – What would be consequences of not using thole dampening in optimization?

    • WW – I tried setting this to 0 in OpenMM and my simulation exploded. And looking at the iAMOEBA implementation (which says it doesn’t use thole dampening) it sets the thole dampening in OpenMM to a nonzero value. Talked to Asimmonett and (see recording ~47 mins)

    • CBy – Sim explosion could be due to overlapping atoms?

    • DM – Re: small vdW on hydroxyl and using nonzero thole dampening factor - This might be coming from the fact that OMM implementation of pol was built for AMOEBA and so they didn’t consider case where dampening is 0.

    • WW – Could try using the HIPPO implementation since their dampening is implemented differently.

    • BS – I’m thinking that the thole dampening factor isn’t important for water. I didn’t expect it to play a role in the penalty funciton.

    • WW – Even when I optimized thole val, the range of values was super small, like 0.390 to 0.400

  • WW – Do you think it’d be a good idea to switch to 14-7 LJ? That’s what AMOEBA uses.

    • DM – Depends on long term goal. To me, the goal is to take a fixed charge FF and add polarization without making other changes to functional form. So I’d view that as a separate project.

    • WW – Agree. Early trials of swapping electrostatics led me to refit vdW terms, and then the next place to look seems to be functional form.

    • DM – I’d recommend trying out nonzero LJ on hydrogens. And I wouldn’t aim at getting the best pure water model, I’d aim at getting mixture properties.

    • WW – Agree. That’s what I’ll do once I have reasonable densities for water.

  •  

  •  

  •  

    •  







Action items

Decisions