2024-10-31 Protein FF meeting note

Participants

  • @Pavan Behara

  • @Chapin Cavender

  • @Anika Friedman

  • @Michael Gilson

  • @David Mobley

  • Louis Smith

  • @Alexandra McIsaac

  • @Jeffrey Wagner

  • @Lily Wang

  • @Brent Westbrook

Goals

  • GB3 NMR fits

Recording

https://drive.google.com/file/d/1S4_2YHNEUBuOcEm7K1c_GFkHJv9i_5Zy/view?usp=sharing

Discussion topics

Item

Presenter

Notes

Item

Presenter

Notes

GB3 NMR fits

 

@Chapin Cavender

  • (slide 6) MS – Should be able to take sims from green line, and since you have predicted NMR params, you should be able to get a predicted FE curve for next step.

    • CC – I think that’s right.

    • MS – Would be interesting to see whether the predicted next step of the green is really the pink line.

    • MS – Other thing could be predicteing the other way - predict confiurational space overlap between two ensembles. Trying to figure out how different what green predicted is from pink.

    • CC – Worth noting that shaded regions are uncertainties from bootstrapping.

    • MS – …

    • CC – Yeah, that could be a good next step. What action would we take from findings from that?

    • MS – If green can’t predict pink then something is wrong. Could be that this process is noisier than we think. Could be that there’s a bug in the process or sampling is happening in different regions.

    • MG – In the limnit of infinite sampling we should have perfect predictions. But may be that we need more and more sampling as we go to lower fraciton of native contacts. CC is running 500ns already, whcih is a lot.

    • LS – Are you doing hamiltonian exchange?

    • CC – No

    • MS – If prediction is bad that’s something I’d suggest.

    • MG – don’t we know that prediction is bad?

    • MS – If green can’t predict pink then we know there’s either a bug or insufficient sampling.

    • MG – If I’m right that the sampling is getting problematic at low % native contacts, maybe we restrict the range a bit? The intiial limit at 0.4 was arbitrary, maybe we could do 0.6… If we had bad sampling at 0.4 then maybe that’s contaminating other stuff?

    • MS – I’m not sure it’d contaminate

    • LS – My experience with doing mechanical denaturation was that coordinate space becomes really large really fast at low nc… Having bad sampling at low nc might screw up…

    • MS – Green curve has samples between 0.8 and 1.0 - it’s suggesting parameter changes to get more sampling between 0.8 and 1.0 and instead it’s getting less.

    • MG – Let mechallenge that - The appearance of what’s happening there depends on the shape of the overall curve - 0 is set at the global minimum of each curve, independent of each other. If we instead aligned at the local minimum over 0.9 it’d look pretty different.

    • MS – Umbrella sims are independent thoguh.

    • LS – Could be good to align them all at the 0.95 minimum that they all seem to have… I do feel like this is a sampling issue. Maybe 1E5 isn’t enough? Or we need a smaller prior so we take smaller steps?

    • MG – On slide 5 we see how little force constants have changed.

    • LS – Good point. That implies that it’s either sampling or a bug.

    • MS – Should be able to directly predict FE curves without any simulation.

    • MG – LS, would it make sense for you to take a look at the low end of these trajectories.?

      • LS – Reason we had those windows was because we had a bunch of sampling in that region (the protien was unfolding). would be interesting to see if it’s folding to some other specific structure or if it’s just disorganized (maybe clustering would tell?). Could be that helices are turning to sheets.

      • CC – I haven’t clustered the umbrella sims. But my initial sims that I used to seed windows - and saw that 1.0 to 0.6 was unraveling helix, and 0.6 to 0.4 is undoing sheets.

      • LS – Another thing could be to look at the all-to-all RMSD between windows (I did that in my original work).

      • LS – Broader Q - Could you try running sims longer and see if curve changes?

      • CC – Can do that

      • MS – I think reanalysis is more useful at this stage. There should be enough info here. FE below 0.6 is probably very noisy, but >=0.7 is probably rich enough.

      • MG – Seems like low-fraction-nc noise would affect high-nc space - it’s pushing on the same parameters.

      • LS – I meant to run another round from scratch, not to extend the current windows.

    •  

      •  

  • MG – Kinda repeating something like LS’s idea - If there’s computer time, it’d be interesting to redo the step after the first-round-fitted-FF again and see if we get the same results.

    • LS – Yeah, kinda

    • MG – I think I’m in favor of “extending” instead of doing a whole “biological replicate”. But ….

    • CC – To be clear - my current thing is where each window is 500ns, and I have 3 for each, and I pool them together for these steps.

    • LS – Also in favor of separating windows.(?)

    • MS – And I’d do 3 preds for 3 sets and see how different they are. We’d hope they’d all be pretty similar.

    • LS – If they were all similar, I’d be convinced that sampling ins’t a problem.

      • MS – Agree

  • MG – Another idea is, given my suspicion that sampling is bad for frames with nc=0.4-0.6, would it be possible to rerun predictions just omitting that?

    • CC – That’s something we could do with the existing trajectories.

    • MG – Right, because doing the splitting of the 3 500ns runs, we’re talking about doing a refit of those and testing 3 different FFs?

    • CC – Idea is just to take the QM fit FF, then to make 3 curves (one with each 500 replica independently). … (see recording?)

    • MS – If you get 3 different results, we know it’ll be a noisy process.

    • MG – Right, and if it’s noisy we could throw out the 0.4-0.6

    • MS – And could do optimization excluding 0.4-0.6.

    •  

    •  

    •  

  •  

  •  

  •  

 

 

 

Action items

Decisions