2020-12-10 Force Field Release meeting notes

10 mins

Brief update on BCC refitting

Simon Boothroyd

• Insert slide link here

• CB – slide 3, bullet 1 – “electric field fitting” –> Was there also ESP used?

  • SB – Just electric field for now, but we can expand to ESP later. And since we store the wavefunction, we can generate out the ESP when we need it.

• CB – Phys prop – Have you looked at hydration free energies?

  • SB – Not in this study, didn’t see it being valuable in earlier tests

  • CB – What if they optimize to different areas of parameter space?

  • (Some discussion)

  • CB – This approach seems like it will come up with an set of BCCs for environments that are less polar than water. DM points out that this will be appropriate for a something like a binding site. But if we want BCCs that are good in water, we should train on some water-based measurements as well, like hydration free energies.

  • SB – We do include enthalpies of mixing with water. Though we could weight those data points differently. I agree that we should TEST HFEs.

  • CB – It may be interesting to see whether BCC gradients for water mixtures are in opposition to gradients from other mixtures.

• CB – I found aldehydes very troublesome in my BCC work. This may be because they expose a carbonyl carbon, which exposes a large face that isn’t shielded by any hydrogens. Expect difficulty with aldehyde hydrogens.

• CB – A potential pathology – Fluorine sucks. BCCs based on QM for fluorine will lead to radii that are unphysically large. On the other hand, restraining the fluorine radius will lead the BCC to shrink a lot.

  • SB – Agree. We’re also really short on fluorine-containing compounds

2 mins

Brief update on fitting experiments

Pavan Behara

• Insert link to slides here

• CB – Good to remember that the torsion fitting and vdW fitting are intercoupled.

• SB – Why do we think fit 3 gets WORSE when non-interpolated parameters are added?

  • PB –

  • JW – Could be arbitrariness of kekulization

  • TG – Could be that the random initial values of the three params for fit3 are too strongly regularized

  • SB – This is a critical table to get right

  • JW – Could explore influence of kekulization by making a new FF, AFTER fit0 completes, with explicit single and double bond parameters for kX_bondorder1 for single bond, and kX_bondorder2 for double bond, and halfway between for the aromatic parameter

• CB – Interested to know why t45 got worse – Just statistical noise, or is it something significant about chemistry?

• HJ – In 1.3.0, there are six sp3-sp3 carbon terms (t43-t48). (I’m not sure I followed this whole point)

  • DM – It looks like fit 3 might have LESS total parameters (…? I’m not sure I (JW) followed this either)

• DLM - should see improvement on 1.3 for interpolation. From non-interpolated to interpolated, the interpolated should make the FF simpler and also more accurate.

• CB - How do we represent adding 1/3 more parameters properly

• DLM - Even if the simpler FF doesn’t perform better, and hopefully we can still argue that fewer parameters is better

• CB - Escaping an aromaticity model is a big win

• SB - How are the WBO being computed? Are they openeye or RDkit?

  • PB - Openeye toolkit

  • SB - Worried we are making an openeye FF, a concern. We want our FF to use opensource tools

  • CB - Adrian Roitberg did a comparison and said that his tool produced the same WBO as the openeye toolkit

  • DLM - Jessica can you compare the WBO with amber tools backend and the openeye backend. Run this one a lot of molecules.

  • CB - Make sure to use single points

  • DLM - recharge a fixed set of conformers. Take a conformer for each molecule and reevaluate the WBO with the amber tools

  • CB - is this the direction of your question

  • SB- Yes. One of them is the underlying AM1 method for a single point calculation work. Also what implementation should we be starting off from? Does the backend make a difference?

  • DLM - coordinate with Jessica offline

  • DLM - When should we set the Sage release? Simon has show improvement. We may want to include WBO interpolation as well. We will need to explore chemical perception of the WBO interpolation parameters more. May need to compare k versus torsion barrier trend. Trends are similar but not same in approaches.

  • CB - how does this affect going forward from this point ?

  • DLM - If we can show that interpolation is not great but improves the FF. We can spend more time on the chemical perception of the FF

  • SB - We should try to show on a more general set the parameters generalize

  • DLM - I agree

  • CB - Pavan needs to finish this direction, then we will have what we need to set a date for the Sage release

  • DLM - Yes, benchmark on not just the training dataset

  • CB - We cannot make a decision the release date unless Pavan’s study finish

  • DLM - We need to make sure we have all the ingredients/double check/polish before setting a release date

  • CB - when we will have the results that drive the decision, question for pavan

  • PB - Not sure, but potentially by next week I will have benchmarked

  • DLM - also wondering why non interpolated is not working

  • CB - Early January

  • PB - yes

  • DLM - The next time we meet for ff-release call

  • CB - Let’s set the sage release date at the next ff-release call

  • HJ - I’ll extend the meeting on the calendar for early January. I am working on the internal coordinate for hessians. I’ll have updates on my project for the next meeting.

  • CB - let’s have you go first next time (HJ)

  • HJ - Still working on the amide fix. Want to test a few more things first

  • DLM - Josh has been looking at bespoke fitting. If you refit torsion for a specific molecule for the 1.0.0 and 1.2.0 release, refitting the torsions could improve energetics a lot. For the 1.3.0 release, fitting the torsion doesn’t help anymore, for one molecule which is an amide

  • JH - will share molecule with HJ

  • CB - should we do a 2 hour meeting or 2 1 hour meetings because we are approaching a critical deadline

  • HJ - will update on the calendar for two hours
    
    

5 mins

Update on amide cis/trans preference fix

Hyesu Jang

https://openforcefield.atlassian.net/l/c/5UKp14yS