Pairwise objective fits  | @Chapin Cavender | [Slides] Pairwise fit uses NAGL charges, best comparison is Null-0.0.3-NAGL vs Null-0.0.3-Pair. Decrease in RMSE on QM test set, and makes alpha helix more favorable in both gas and condensed phase Unfortunately doesn’t keep helix during simulation Slightly lower x2 values on NMR data, but large error bars
Good baseline for new objective function, will modify fitting procedure to try to improve DM: curious how this interacts with reweighting from previous meeting, but that’s the next topic JW--agree with these conclusions. do you have QM data for spurious MM? CC--no, need to generate, but can do locally JW--QCArchive is wide open so can do it there if you want. Probably will want it there eventually
|
Quick validation by reweighting | @Chapin Cavender | Goal is to triage bad force fields before doing a full benchmark, since benchmark is expensive Reweighting works well if there is a lot of overlap between the query and new, but there isn’t much overlap between ff14SB and SMIRNOFF, so want a new trajectory JW--Slide 16--thought you had said if distances are all perfect, arg of exp becomes 0, if Rij = R0ij. Shouldn’t the perfect value be 1/2? CC--yes, it’s like a sigmoid function. If you are far from native contact, you contribute 0, if you are perfect you’ll contribute 1/2, if you go past you’ll contribute more JW--ok, but shouldn’t a perfect score be 1/2? CC--contact width parameter is 1.8, so score would be 1/2 at 1.8 x the native distance, not when it equals the native distance JW--so it would be happiest if everything was below the native distance. wouldn’t it be happiest if the protein collapsed into a black hole? LS--wouldn’t you get a positive number if you did that? [some discussion around 25 mins] JW--OK so higher score is better? CC-- yes, we’re assuming the FF is good at handling things not collapsing into a black hole
LS--Slide 18, can you remind me what the baseline was for these results? Was the trend similar to the RMSD? CC--Yes, Null 0.0.3 has low/stable RMSD in OPC, specific is worse LS--Could we just use this trajectory (Null in OPC) for reweighting? MG--would expect better reweighting for results coming from OPC. If we were happy with just using OPC, we could have been done a while ago AF--I have simulations related to this and can present them at the next meeting. Some targets were decent, some really good, only significant difference was lysine ( ? ), still looking into issues CC-- will look into it
DM – will be most important to take unfolded Null and get it back to folded; there will be infinite ways for it to unfold, and many of them won’t have any overlap still CC--yes, trying to generate sample trajectory that goes between eg 0.3 and 1 DM – Two things you’ll get most value of is taking 0.3 and getting it back to all native contacts, and starting with FF14SB and changing the force field but keeping it at all native contacts (could be artificially restrained) CC – ok
JW-- wanted to double check on null 0.0.3 with OPC--what I recall is that things that looked good on 10-mer looked bad on GB3 and vice versa, did this do the same? CC--yes, looks bad on 10-mer CC--basically back now at the performance of the original Null 0.0.3, but can fit in 2 days instead of 2 weeks
Discussion between DM, CC, and LS about details of umbrella sampling around 40 mins
|
