2021-01-06 BespokeFit meeting notes

Planned methods for torsion fitting

• DM – (Recounting previous findings on doing MM optimizations during fitting to torsion drives – Basically, “should we be optimizing more than we already do?”)

• DC – Instead of restraining less than we already do, should the restraints be built into the ForceBalance objective function?

• JH – My question would be, removing …?

• DC – Recap

  • Question – Which torsion protocol should be used in QubeKit? Decided we want to try doing things using OpenFF infrastructure, but with modifications/extensions.

  • Result – JH will run torsiondrives using traditionally problematic molecules to see if we get the MM and QM barriers spot-on. If we can find some problematic cases, we’ll look at possible changes to the objective function that consider the RMSD between QM and MM final structures. Focus on timing differences between doing full MM optimization vs more limited approaches.

Recap of results of torsion fitting

• JH – Have most of the data, need to compile a full summary, will update when that’s ready for discussion.

Fitting using modified seminario method?

• JH – Might expect this to avoid some problems that we’ve seen that require bugfixes (eg. between 1.2.0 and 1.2.1, problem with 1.3.0)

  • DM – The poor current performance might be due to this particular test set. It might be worth looking at the different parameters that are being applied, and seeing the effect of mixing parameters into the final system from GAFF or modified seminario-derived parameters.

  • DC – CR has looked at pulling large numbers of hessians down from QCA and seeing how well the modified seminario method does. However the data is quite complex and will require more analysis.

  • DM – Shared figure from Parsley preprint – “Even if things seem hard to split up, it may be good to look at how broad the distributions are”

    Open

    

  • DC – It may be informative to do this sort of analysis based on which parameters are assigned.

  •  

• JA – We keep coming back to the 1-4 scaling of vdW and charge interactions. This is something that we can’t really change in the general force field, but we may be able to safely do this for bespoke torsions. It may also be possible to have per-torsion 1-4 scaling factors, though this would be more difficult in current MD engines.

  • DC – We were thinking about doing this with the LJ potential in the future. Though I’m not sure how we’d decide on which scaling factors to use. We’d need to find some very particular torsions to find cases where we could test this.

  • JH – Maybe some of the molecules that were found in the bond order interpolation work? I know we’re trying to make a dataset of sterically congested torsions.

  • JA – Or protein backbones could be good for this.

  • DM – Changing 1-4 scaling factors is on on roadmap as well, but we haven’t begun work on it.

• JA – Does anyone have experience with fitting torsions for things like cyclohexane?

  • DM – Most of that expertise, if anyone in OpenFF has it, would be in Lee Ping Wang’s group.

  • DC – We’re going to start looking into this soon.