Evaluator issues | BM / MS | MS: the water 2ns simulations are consistent in the sense that the variations are on the scale of the error bars, unlike “component 1” (triethanolamine) LW – Slide 10 – Is this production or equil sims? MS – Spikes in energy are causing it to only use 500-1k data points. DM – What’s the thermostat? MS – Should be langevin middle. Not sure how this would happen. Spikes don’t follow boltzmann distribution … MS – With OpenMM, it’s hard to tell what it’s doing. One thing on our to-do list is to run these with default settings in OpenMM and compare to GROMACS.
MS – Not clear why it takes some mols a nanosecond to equilibrate (but water equilibrated fast). Also appears that equilibration detection is broken. Also not clear why there are huge negative energy spikes. PB – Do we have data from earlier fits? CC – Do you know whether energy or pressure*volume terms are causing spikes? MS – I think it’s likely to be energy - For this system PV should be negligible. And kinetic energy must be a gaussian at this scale. BM, could you do QQ analysis for KE and PE separately? DM – Do you see this for other compositions? MS – Slide 5 and 6 is particularly confusing. CC – You were still seeing spikes in enthalpy with pure water? MS – A little, but not nearly as much. Water data mostly looks spiky because of narrow yscale. CC – If we’re not maintaining target temperature then ??
MS – I think big spikes are causing equilibration detection to be bad. JW – Anything without hydroxyl Hs in solute? Maybe try increasing H radius. Could there be collisions? DM – I’d try taking polar H sigma and make it as big as normal H sigma. My expeirence is that polar H LJ radii haven’t gotten much attention, and maybe folks hadn’t tried doing mixture sims/FE sims that would have revealed these issues. MS – Then why is normal TIP3P (with 0 H radius) so well behaved?
CC – MAybe try increasing collision rate on langevin thermostat? Could help control temperature better. MS – What’s collision freq on thermostat? LW – Recommended default is 1/25ps. MS – That’s kinda long, should be 1/2ps. Barostat changes every 25 steps. LW – My mistake, openmm docs recommend 1/1ps CC – Can serialize OpenMM system to XML MS – Is there a way to do that? LW – Not easily, it serializes State sometimes, but not System. DM – If science team has bandwidth this would be good to look at. LW – Yes, this seems pretty important, also for current sage refit. Also, I’m working on some stuff with OpenFE to do… (see recording ~31 mins)
BM will do QQ plot of PE and KE BM will try running with larger sigmas - Can reach out to LW for assistance
|
NAGL | LW | CC – When you say result is identical to numpy preision, there’s no change? JW – 2 Qs: LW – Make this be the full release? JR – Does this casue problems wiht divalent vs. trivalent ions? Like iron? MS – Does this currently handle sodium chloride? Will current standard workflow handle sodium chloride? JW – “full release” means transitioning from phase 2 to phase 3? NAGL fingerprinting tool discussion - See recording ~55 mins LW – What should ESP RMSE cutoff be? CC – A molecule could belong to multiple groups? PB – I’m still on the fence - This looks like a great tool for diagnosing and correcting things. But I don’t see why someone running an FEP calc check DM – A bunch of people have asked for it - including funders and openfe. LM – I think people are also skeptical of ML. PB – I guess I’m a little concerned that we’re setting a high standard and this will set higher requirements for future work. LW – Yeah, that makes sense. For AT and OE I trust them because there’s so much mileage on them. But we’ll need to build new trust for the ML model.
LW – So we’ll continue with releasing NAGL into phase 3 as the default backend, but not have substructure tool used by default. ……… Primary decision on whether to advance nagl from phase 2 (which we’re in currently) to phase 3 (installed by default and used as the default AM1BCC provider) will be next week.
|
