Page Comparison

Relative conformer energies in the public dataset

Bill Swope, Alberto Gobbi, Ben Sellers

• https://docs.google.com/presentation/d/18PvATGQ7ovhBUF4QlAUMA2EaprIr8Wvp91y6uMuXzwA/edit?usp=sharing

• LW – happy to re-do scripts in Python, or work with Bill to do so

• DM: are you happy to look more into this and identify specific functional groups, problematic chemistries, etc.? e.g. using tools like checkmol

  • DM: might bring this up with Lorenzo

  • WS: e.g. splitting broad parameters into multiple specific ones?

• WS: what goes into the forcebalance training objectives? Do you have multiple conformers per molecule?

  • PB: yes

  • WS: and using the Hessian?

  • PB: not in Sage – included in Parsley

• WS: Alberto noticed problems with ring conformers, thought it might have something to do with overly stiff angle parameters

  • DM: ring parameters have some problems – philosophy was to let geometry sort it out, which results in alright structures, but energy issues

  • DM: would be helpful to pinpoint if that was an issue, ring parameters haven’t had a lot of attention so far, could be more specific

• WS: different problem molecules seem to have different interactions/problems

  • DM: Lorenzo’s been looking at overrepresented parameters in problem molecules – should probably repeat on molecules identified by Genentech’s analysis

  • DM: typically bad params fall into two camps:

    • too broad, or

    • trained to different chemistry to which it’s used

• WS: happy to work with OpenFF and keep helping

  • DM: will work out good pathways internally

• General – get back to us in a month