search

choderalab / gbff

Tools for Bayesian forcefield development
3 stars 6 forks source link

RMSE vs iteration #19

Open jchodera opened 9 years ago

jchodera commented 9 years ago

I'm looking at the RMSE (in kcal/mol?) vs iteration plot from the 300_adaptive_3gbmodel_largejoint_days.h5 dataset.

Any idea why the initial RMSE would be ~13 kcal/mol? What parameter set was the model initialized with?

RMSE.pdf

jchodera commented 9 years ago

Here is RMSE and sigma, which hovers around 6 kcal/mol.

rmse-and-sigma

pgrinaway commented 9 years ago

looking into this now...

pgrinaway commented 9 years ago

I think this is a unit issue. The returned energies were converted to kJ/mol and the measured energies were kcal/mol. I ran another run w/ the correct units last night if it is of interest.

jchodera commented 9 years ago

Good catch! Would love to check the data --- where can I find it?

Did you check in the fix? Apologies if I just missed the PR.

Thanks!

pgrinaway commented 9 years ago

Unfortunately it doesn't have too many iterations but it's available at /cbio/jclab/home/pgrinaway/gbff/300_rightunits.h5. RMSE appears to go from ~6.8 kcal/mol down to ~3.9 kcal/mol. I had just checked it in to my own repo--pull request filed.

jchodera commented 9 years ago

I'm still confused as to why the RMSE is so high initially. 6.8 kcal/mol still sounds way too high---what model is used initially? The OBC GBSA parameters?

jchodera commented 9 years ago

Hm, I wonder if we're changing the SASA penalty term as well in the fitting procedure. I totally forgot about that, I think.

pgrinaway commented 9 years ago

OBC is used here initially. And no, I don't think we are changing the SASA penalty term.

jchodera commented 9 years ago

Thanks! I've opened #22 to remind me to fix that.

jchodera commented 9 years ago

It looks like David Mobley hasn't benchmarked FreeSolv with either implicit solvent single point or free energies yet. I can write a small script to do this with OpenMM but worry that there may be nontrivial differences with Amber implementations of the models. Was there a way to use parmed to compute these GBSA energies?

kyleabeauchamp commented 9 years ago

So that was something that was done using pytraj, discussed in this thread: https://github.com/choderalab/gbff/issues/8

However, the IPython notebook link is now dead, so you'll have to dig around for the latest link.

On Mon, Oct 12, 2015 at 4:36 PM, John Chodera notifications@github.com wrote:

It looks like David Mobley hasn't benchmarked FreeSolv with either implicit solvent single point or free energies yet. I can write a small script to do this with OpenMM but worry that there may be nontrivial differences with Amber implementations of the models. Was there a way to use parmed to compute these GBSA energies?

— Reply to this email directly or view it on GitHub https://github.com/choderalab/gbff/issues/19#issuecomment-147549210.