2020-06-17 BespokeFit meeting notes

Bespoke workflow planning

• How do we get people to select dihedrals?

  • JH – Torsiondrive uses text files

  • JW – Could introduce a plugin architecture, where one solution just reads an input text file, but future ones could be more sophisticated, or open nglview and let people select bonds.

  • HX – We take approach of accumulating data, where we record all of the fits, and we’ll jsut load previous results of people select a torsion that’s similar to one that we’ve fit before.

  • DM – We could accumulate “goodness of fit” stats for different parameter types, and use that to decide whether we need to refit it or not.

  • JW – So, a torsionselector plugin, where we could start with EXHAUSTIVE, but move on to MANUAL or variants of SMART (which tries to decide which torsions are likely to be inaccurate)

• JH – How accurate do we want generated data to be? B3LYP will be quite slow on large molecules or fragments.

  • JW – Per previous discussion with Chodera+Horton, ANI would be best

  • DM – Yes, B3LYP would come later, and we shouldn’t worry about comparisons initially.

  • JA – We should focus on accuracy first, so I’d be in favor of prioritizing the use case with B3LYP first.

  • DM – For a lot of use cases, speed will be important, so we’ll eventually want to support both, but ANI will be a big use case.

  • JA – What use cases do we expect for this workflow in pharma?

  • DM – We see pharma doing this on something like 50-100 molecules, in preparation for various types of workflows.

  • HX – Normally, we do this type of calculations for about 300 molecules/week. FF fitting and FE calcs take about equal amounts of compute

  • JH – which basis?

  • HX – Use standard HF for electrostatics, different levels of DFT for geom opt and PE surface scan. Working with MolSSI to benchmark DFT levels. No conclusions on this yet, we’re busy with some other things.

  • DM – Please keep us posted.

  • HX – Are you considering using ANI to replace QM calculations?

    • DM – We can use it in the same workflow, as if it were a QM calculation. So it’s handy for debugging. And we’re interested in benchmarking it for torsion parameterization.

    • DM – (shows Rowley benchmark comparing torsion drives)

    • JA – Is ANI trained on biaryl rings?

    • DM – It will have seen some, but not all.

• JW – Will volunteer josh the be “in the line of fire” for future submissions (Mostly Cerutti/protein fragments)

  • This will involve making package releases of QCSubmit, but keep in mind that this is NOT expected to be a stable release. Just a way to go back and reproduce submissions

• DM – Is there anything close to our current path that could justify Silicon supporting us?

  • HX – Not currently, but will think about it.

• JW – Let’s see if we can hand off a MWE of distributed QCF job for Galileo team in 3 weeks

• JW – JA, we’re getting you on confluence

  • will do this immediately following meeting

• JW – HX, send QM candidates your way?

  • Yes