GB3 | @Chapin Cavender | CC – summary of issues with previous SMIRNOFF simulations – alpha helix unfolds from C-terminus in at least one replicate CC – currently running simulations of short alpha-helical peptides with amide cap. This required deriving LibraryCharges for the amide cap, same procedure as I used for others. I get charges similar to previous values, and they only differ in 4th decimal place, so probably close enough without repeating with 5-mers MS – why was this necessary? CC – molecules in NMR experiments have caps, and other FFs support this, so seemed like we should support it anyway
CC – simulations with AMBER can sample helical conformations within 1 us, continuing them to hopefully see multiple un/folding events CC – hoping to show results at next meeting CC – trying the Ramachandran clusters of secondary structures of unstructured short peptides doesn’t seem to show the short peptides to sample the alpha helical region much. MS – using AMBER or SMIRNOFF? CC – AMBER is also not helical (shows slide) MS – but there are differences between them CC – basins (beta, P_II) favoured by SMIRNOFF vs AMBER respectively are close to each other on the map CC – alpha helix basin is subset of delta basin for ala3 CC – delta basin is sampled a bit more by AMBER but only by a few % CC – we see more differences in Ramachandran clusters for gly3 CC – repeated analysis for GB3
CC – looking at side-chains with same Ramachandran analysis now (slide 63 – peptide GVG) t = trans, m = gauche -60, p = gauche +60 in PDB, rotamers are almost always trans, but much less so in simulation. In simulation, rotamers are often at angles that fall outside the t/m/p bins See similar behaviour in val3 and gb3 too Looking at methionines, we get more complexity
CC – NMR observables for GB3 AMBER models them better than SMIRNOFF. Error bars from bootstrapping MS – so Null is doing better than Specific? MS – but we do comparably or better on smaller peptides? Slide 74 CC explains key of x-axis. nh == amide, cb == beta carbon, cg == gamma carbon, etc. In general AMBER models side-chains and H-bonds better CC – Both FFs struggle with H-bonds (last column), but SMIRNOFF does about twice as badly MG – is this analysis telling us somethign we didn’t know? If we took out the problematic helices, would these numbers settle down? CC – can do this experiment easily, but what this is telling us already is what the differences are between the two force fields. When we compare backbones, we don’t see much difference, so they’re probably not the problem. The problem is likely in the side-chains and H-bonds. MG – the unfolding is the result of many cooperative events, and it’s hard to know how to weight what. MS – correlation vs causation – if it’s started unfolding, the H-bonding will be worse MG – Yes, it’s hard to know what to pinpoint here. CC – Agree, will repeat this analysis without the unfolding helix. But we only see the unfolding happen in one replicate out of 3, and it takes a µs to get there. We’re sampling this unfolded state at worst, a third of the time. But the errors are 2x worse for SMIRNOFF CC – will repeat this for 2 replicates without unfolding
CC – my takeaway: we should focus on investigating side-chains instead of backbone DM – what are good focused experiments we could do to isolate where things are going wrong? We would need a cheaper test than simulating the whole system. DM – for nonbonded errors, if we’re using LibraryCharges, that means problems would be from our LJ refit, right CC – could also be from the charges. Comparing ours to AMBER’s, ours seem larger for polar groups and smaller for non-polar MS – that would make H-bonds stronger, not weaker DM – should we swap in the AMBER charges and see what happens? MG – are you suggesting a PMF or another GB3 simulation? (after some discussion) GB3. Probably will take a week or two, it’ll take little of my time to set it up while I do other things
MS – it would be more interesting to look at deviations when it’s folded than unfolded, because that’s causing the unfolding. MS – going back to the BB analysis, there are clear differences in how AMBER does better on gly/val and SMIRNOFF on bulky hydrophobic residues MG – did the SMIRNOFF ffs have higher RMSDs than AMBER? LJ parameters might be of interest, swapping them out for AMBER values should be straightforward
MG – if we need to tweak parameters, do we propagate the changes back to small-molecule world? MS – yes MG – proteins might be more sensitive to subtleties, if they’re near conformational change MS – most fully folded proteins are stable 10-15 kcal, or at least more than 0.5 kcal MS, MG – historically protein FFs have been overstabilised – it’s better now, but just because we have a simulation that’s more stable and folded, we can’t necessarily say this is better without other evidence MS – it’s more problematic that we do worse on experimental observables
|
Benchmarking resources | @Chapin Cavender | CC – maxed out of computing resources. I can share my scripts if other people can run these DM – are these benchmarks ready to go? are they already set up? CC – main unknown is what starting structure should be used, what solvent conditions should be used. There’s still some work to figure out what temperature the NMR was done, etc. But I can look into that and work them out in a single work day MS – one possibility: Anika could do this in our group. MS – can we use non-OpenMM engines? e.g. gromacs. CC – can share my code (which was intended for release in the future) and point out where you can use gmx instead of openmm, assuming interchange supports it MT – I’m pretty confident in interchange, not sure everyone else should be. They’re ready for use - make sure this is using 0.3.0 or newer MS – suggests comparing energies at time 0 between OpenMM and GMX as validity check
MS – can get 3 million hours in a week of time. Have you asked for time via ACCESS yet? You can now write proposals per grant and get up to 1.5 million units (1 unit ~ 1 cpu hour). Turnaround time for grant is a day. You can write a proposal for each grant, so if you have multiple grants supporting the project, you can have multiple proposals
DM – if Anika wants to, benchmarking more systems would be beneficial (general agreement) MS – CC and AF and I can meet to figure out allocations of work
|
