jchodera
commented
4 years ago I've you're using an ML framework anyway, why not just train a more powerful graph convolutional model that you could fit along with your other parameters? See our example.
Closed proteneer closed 4 years ago
jchodera
commented
4 years ago I've you're using an ML framework anyway, why not just train a more powerful graph convolutional model that you could fit along with your other parameters? See our example.
proteneer
commented
4 years ago What’s the size of parameter space in your graphical models?
On Tue, Mar 3, 2020 at 5:39 PM John Chodera notifications@github.com wrote:
I've you're using an ML framework anyway, why not just train a more powerful graph convolutional model that you could fit along with your other parameters? See our example http://www.choderalab.org/publications/2019/9/17/graph-nets-for-partial-charge-prediction .
— You are receiving this because you were assigned. Reply to this email directly, view it on GitHub https://github.com/proteneer/timemachine/issues/158?email_source=notifications&email_token=AARM2FAEVVJO5FTTAMF566TRFWBJDA5CNFSM4LAU4EF2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOENVOKYA#issuecomment-594208096, or unsubscribe https://github.com/notifications/unsubscribe-auth/AARM2FCFLDQCKTUFDFWEN4TRFWBJDANCNFSM4LAU4EFQ .
-- Yutong Zhao
jchodera
commented
4 years ago What’s the size of parameter space in your graphical models?
@yuanqing-wang and @maxentile can comment on the technical details!
chuantian
commented
4 years ago In the paper, the DFT was computed with dielectric permittivity =4 (I suppose doing so to make the resulted charge in between gas-phase and solvent-phase), which reminds me how charges in ff03 amber force field (Duan et al.) were trained. This might suggest a reasonable combination of GraphNet with ff03? @jchodera
yuanqing-wang
commented
4 years ago What’s the size of parameter space in your graphical models?
The number of tunable weights? Around a few thousand would do.
proteneer
commented
4 years ago We're now going to directly port over the excellent patterns being developed by @SimonBoothroyd as part of the openforcefield recharge project.
proteneer
commented
4 years ago This first pass is done.
I mean, they're called bond charge corrections already right? So we might as well as just try to use the bond charge information.
87 parameters in total
One disadvantage of this method though is that we won't be able to lower the norm of the avg charge.