Page Comparison

Updates

• MT

  • VSites

    • Gradually adding in functionality, yet to add an “everything bagel” molecule

    • ChargeIncrements are a bit complicated, since they seem to have different meanings in CIMH and VSiteHandler

      • CIMH just adds the increment to each tagged atom

      • VSiteHandler takes some positive/negative charge from the atoms, and put them onto a non-atom

    • Worked with TG on how our vsites interact with meaning of FD/2FD/3FD – These largely differ by how they define distance (either explicitly, or relative to some other measurement). I’d picked the wrong one initially, but now I’m on the right track.

    • Found a handful of bugs with AMBER energies. Not sure whether these are in the file outputs themselves, or sander settings.

      • Github link macro
        link https://github.com/openforcefield/openff-interchange/issues/249

  • Two toolkit bugs

    • One was a simulation crash bug, was due to using different HMR settings

      • JW – Is hmass=4 a normal thing to do? This makes methyl carbons really light, and based on my visualizaiton as the root cause of the crashes. Will need to either add new canary tests or declare that we don’t support hmass=4

        • MT: 4 amu is in openMM tests, so at least can’t be totally outrageoushttps://github.com/openmm/openmm/search?q=hydrogenMass

        • LW: Some cases in which 3 amu is used in Amber (GPU reasons?) Improving efficiency of large time‐scale molecular dynamics simulations of hydrogen‐rich systems

    • OFFTK #1019 highlighted an error in ParmEd use, which garbled 1-4 energies.

• PB

  • Worked on theory benchmark, dataset submission and analysis scripts

  • Tried to reproduce TF’s problem, no success.

    • JW – Let’s see what he says, right now we don’t have the info we need to reproduce

  • Re: Chapin meeting – Will we need to do 2d torsiondrives?

    • JW – I’d expect that we might need to do this eventually, but it would be reasonable to restrict to 1D initially, so we can do a cheap first iteration before we commit a lot of resources

    • SB – We have some 2D torsiondrives going, but they’re proper+improper. They’re also taking a really long time, seems like a lot of dead jobs. Not sure what’s happening here, may be worth investigating.

    • JW – Could look at the root cause for these – If it’s job pre-emption on the compute clusters, we might consider just spending money on dedicated compute.

• JW

  • Vacation for a lot of past two weeks

  • ParmEd/OFFTK #1019

  • Benchmarking release w/ Dotson.

    • Next release needs

      • sage rc1 to be pacakges/available (done, with new openforcefields pacakge)

      • new analysis method (LD is working on this)

    • LD

• • • found that sage rc1 has a high error rate in MM minimizations, though it could be due to compute resources. DD is working on reproducing with public dataset

      • SB – I ran through the public set using sage rc1 before the workshop and didn’t see a huge error rate

      • JW – There may have been a bug in the error reporting, so I’ll check with Dotson about this and get back to you Wednesday.

  • Next up

    • Finish debugging OFFTK #1019 and update notebook

    • Announce new FFs (1.3.1 and 2.0.0-rc1)

    • Followup workshops announcement

    • Fox analysis debugging

    • Benchmarking release

    • 0.

• • • 10.

• • • 0 OFFTK release

    • 2021 Roadmap update

  • SB – How do pharma partners feel about docker?

    • JW – I can’t recall if we polled them about this.

    • SB – This would be a way easier release pathway. The benchmarking+bespoke stuff could have a much easier time deploying internally+externally.

• LW

  • Was off most of the last few days

  • Read CC’s plan for protein FF. Tried making a “null” FF, where I make a protein FF using the amino acid building blocks.

  • Tried using Psiresp to generate charges.

  • OpenEye wrote back to David about atom-order-dependent geometries (and therefore partial charges). They told us to do our own minimization, and when I tried this out (using RDKit to minimize) I got much better charge consistency for different atom orders.

    • SB – Have you tried completely disabling AM1 geometry optimization?

    • LW – Yes, this gives consistent charges.

    • JW – Connor Davel (Boulder undergrad) has shown preliminary data that shows that charges don’t really change if we don’t allow geometry minimization at all

    • SB – It’d be good to check in with CBy to see if there are edge cases where minimization is really important.

    • LW – What kinds of molecules is CD working with?

      • JW – 20ish heavy atoms

    • SB – Has anyone tried using PySCF? They seem to have an AM1 implementation

    • Github link macro
      link https://github.com/pyscf/semiempirical

• SB –

  • Out on vacation last week

  • Opened PR to QCEngine to do torsiondrives. MolSSI/Dotson will need to review/decide on whether this fits with their architecture

  • qcsubmit is on conda-forge! Now compatible with new namespaces. (fragmenter also got added around the same time)

  • Working on getting bespokefit ready for september workshop – Additional pedantic-friendliness, using fastapi, more microservice-focused architecture.

Sprint planning

Meet at 46 past the hour.