Page Comparison

GB3 NMR fits

Chapin Cavender

• Slide 5

  • MG – The bottom row don’t look like replicates. Suggests that sampling is insufficient.

• Slide 6

  • MS – One thing that may hurt sampling is that there’s no replica exchange, and that umbrella sampling on uncorrelated samples is (bad), since ex two confs with same native contacts might not be able to transition to each other without leaving umbrella window. So not surprising to see … So my suggestion is to do replica exchange, though I know that’s a lot of work.

    • CC – In December I was thinking of doing a cumulative fit, where ex I fit to green and orange line.

  • JC – What is metric of uncertainty?

    • CC – SEM from bootstrapping over samples from MBAR. So I go through the process of identifying uncorrelated samples using pymbar. Then I bootstrap over the samples to get the new FE curve, and I sample over that… to get the error.

• (Slide 8): MS -- the resampling involved a new simulation?

  • CC: yes

  • MS: so you need to do additional energy evaluations over the pooled dataset?

  • CC – Right. I take the trajecotry of the orange line, and compute the energy difference between orange and green… (see recording ~18 mins)

  • MS: that process shouldn’t make it worse

  • CC: I agree

• MS: are you doing an optimization with the pooled data?

  • CC: yes

  • MS: the decrease in performance with the addition of more data doesn’t make sense, unless the initial optimization was just lucky in how well it did.

  • CC – Lucky in what sense?

  • MS – That it got better on data that it didn’t actually see. When you add new configurations, all the reweightings should be… like, how can it be so wrong about configurations it’s seen before?

  • CC – I think this is explained by your thing before - that there are many distinct confs with the same fraction of native contacts.

  • (lots of talking about lines in figure, see recording ~24 mins)

  • LS – In butane experiments, you have to iterate twice to get cumulative trajs to reproduce the issue. So it might be interesting to do that with intentionally insufficient sampling, since that might reproduce some of the issues that we think may be at play re: insufficient sampling in GB3.

  • MS – My current best guess is that it’s a software bug. So I recommend trying butane to see if there’s a software bug,

  • MG – Yeah, so try taking another step WITHOUT using cumulative traj?

  • CC + MS – Yes

• LS – Maybe take a look at frames from null vs. other sim to verify that they’re really finding new minima?

  • CC – Yeah, good idea.