Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

View file
nameFFR-2023-06-22.pdf

\uD83D\uDDD3 Date

\uD83D\uDC65 Participants

\uD83D\uDDE3 Discussion topics

Item

Notes

GNN update

Slides will be uploaded

Link to recording: https://drive.google.com/file/d/1yCFviqTtkomfB4jXilw2jSxofIQBFQbX/view?usp=drive_link

Slide 2

  • CC – Differnce between “denominator” and “weight”?

    • LW – They’re not very different - The error is multiplied by the weight and divided by the denominator. So they could be one number if we wanted.

  • DM + JW – The GNN_reg models look really good. Basically clearly better than AmberTools.

  • MG – I’d like to check that molecules with like carboxylates and methyl hydrogens (symmetric groups where charges should be identical) are actually identical

    • LW – I’m pretty sure this is the case. I’ve hand-checked a few systematic symmetries (e.g. carboxylates) for symmetric charges and have found that to be the case.

  • JW – Is there a way to tease apart whether it’s a problem with (small errors on high-absolute-value charge atoms) versus (charge transfer long distances across molecules)?

    • LW – Kinda, I could look at this

  • CC – It seems like including physical properties in the training could make sense

    • LW – That’s what I’m leaning toward.

✅ Action items

⤴ Decisions

...