Page Comparison

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.