2020-07-15 BespokeFit meeting notes

QCSubmit validation

• JH –

• Dataset validation has been implemented in a GH bot

• JW – QCSubmit should include its own version number in checks, as well as OETK’s and RDKit’s versions. This info could be in a collapsed field in the markdown.

•  

Bespoke workflow + QCSubmit

• FittingSchema is coming along nicely

  • JW – Should check in with LPW/HJ about fitness of this schema for ForceBalance

• Some trouble with symmetric molecules and not understanding that two torsions can be equivalent.

  • JW – We could add ChemicalEnvironment.__eq__ to OFFTK that indicates whether two SMARTS are identical. But this will take a while. So for now just implement whatever you can.

  • DM – Maybe chemper would work well for this?

  •  

• JH – Bespoke workflow should submit to a snowflake server. I’ve built in an option where you can turna. FittingSchema into a QCSubmit input, and then you can load the results into the schema.

• JH – General class for targets, which can easily be extended with new types. For the ForceBalance-related targets, this will do the conversion to graph molecule (currently mol2)

  • JW – Should also make it output SDF

• JH – ANI consistently crashes when running locally. ANI2/TorchANI2 may be more reliable.

• DM – This is interesting because we’re benchmarking against OPLS3e, which sometimes does bespoke fitting. This dataset has been approved for release, so it may be good for our benchmarking.

• DM – Could bring this up at ad board in a few weeks. They’ll be interested to see it.