Page Comparison

GB3 NMR fits

Chapin Cavender

• (Discussion of slide 2)

  • JW – (generally missing the point)

  • MS: takeaway is that the dotted green and solid green lines should overlap, happy to take a look at the code and help debug.

Slide 8

• MS – Feasibility of doing umbrella sampling on native contacts? Umbrella sampling on RMSD tends to slow things down a lot becasue the space is so large.

  • CC – There are fewer paths to do that if we focus on just helical residues.

  • MS – But then we might be overfitting.

• AF – We’ve seen loss of beta sheets as well, but not to the same extent as alpha helices.

  • CC – From what I recall, the beta part is two stacked hairpins, and the main feature is that they fray, and that motion is consistent with the NMR data. So that’s something we should sample and the NMR data can tell us how much. But with alpha helix loss that’s not covered in the NMR data, and what follows is that the beta sheets nearby are also lost.

  • MS – Since these are running in OpenMM, what is preventing us from doing repex?

  • CC – It’s a solvable problem, but will require additional scripting.

  • MS – Will improve sampling, but not sure how much.

• MS – Other possibility is that the changes in parameters are too small. It’s a tradeoff between noise and …

  • CC – I use log_10(alpha) = 4…

  • MS – Could be try doing something between 3 and 4… It’s more likely than not that you’ll see significant differences since you’ll be taking bigger steps. The big question there will be whether the predicted curves end up being more accurate. But still worth working on the cumulative stuff too. Will reduce noise, increase n_eff samples, and likely allow you to take biger steps.

• CC – So I’ll keep the additional iterations working, and will focus on debugging cumulative fits.

Protein QM fit experiments

Lily Wang

DM (chat) – I wonder if there’s a way to tell WHICH small molecule data wants to drive it that direction (away from AMBER)… like, maybe it’s different chemistry, or weird sterics causing problems, or…

• CC (chat) – Does the magnitude of the gradient per target at step 0 of the ForceBalance optimization tell us this?

Slide 16

• AF How different do these torsions look when they appear in small molecules?

• LW – Great Q. But worth mentioning that the olive/brown curve started from the amber values and stayed relatively close.

• CC – What’s being plotted here is the raw dihedral term, not the total torsion energy.

• LW – Right, next step is to plot total energy profiles for torsiondrives of these

• CC – Can share scripts on how to do this.

Slide 17

• MS – FF14SB doesn’t have sequence specific phi and psi, so why is phenylalanine barrier different?

  • CC – The way that AMBER14 solves this is that they have a general parameter that goes into all backbones, then a special one that goes into non-glycine residues

  • MS – That peak seems really big

  • LW – Possible that Amber overstabilizes proteins. But the fact that the shapes are different might mean a lot.

• MS – Is it clear that there’s a “right” answer that we can use for comparison?

  • LW – …

  • …

  • CC – Could compare to training torsiondrives. But generally the SMIRNOFF FFs are closer to QM. I think the big bump in the ff14sb profile in this torsiondrive comes from just QM fitting.

  • MS – …

  • LW – Will plot against torsiondrive total energy.

Slide 23

• JW – The y offset maybe makes this look worse than it needs to

Protein QM datasets

Anika Friedman

• JW – Are these planned for submission as opts or torsiondrives?

  • (general) constrained opts

• MS – Would be good to compare dataset composition to what amber used for fitting.

• CC – This looks good, this is what i was hoping to see, and the distribution on the ramachandram map

• AF: can I talk to you, JW, about getting these submitted?

  • JW: JC + LW are the peopl

  • AF – This will be my first submission, could use pointers.

  • LW – Want to join next week’s QC submission meeting?

  • AF – Yes, that’d be great.