General updates | -- Dear Diego, Perfect! This report now has everything in it I expected. It also highlights the accomplishments and also explains the two branches โ industry funded and NIH/Danny funded. This will not change our decision to direct our funding to OpenFE, but hopefully it will further make other companies more willing to contribute to OpenFF. -- At the last Advisory Board meeting it was clear that there is no interest in the release of a new version of Sage. Gary Tresadern said it wouldn't help him make the case with his managers in favor of keeping OpenFF funding and both Chris Bayly and Charles Hutchins said it wouldn't make a difference to OpenEye and AbbVie respectively either. In the same Advisory Board meeting, it was pointed out that the most interesting thing would be for OpenFF to be willing to address the "gaps" that may exist with regard to force fields. Because of this, last Thursday's ff-release meeting was entirely dedicated to the discussion about how would be the best format both to collect information about what these gaps would be and how to attack each one of them. The proposal is to carry out a "microworkshop" focused on gathering information about the "big failures" and "greatest desires" of our industrial partners regarding the use of force fields. I asked for suggestions on the #ff-release channel and will bring this topic to the agenda for tonight's lead team meeting.
|
Individual updates | MT Found nglview is broken - Upstream had a major version bump. I reported it to Hai but didnโt hear back. The thing to do is pin ipywidgets<8. Lost a bit of time last week to jax/jaxlib packaging confusion. Seems to have different maintainer every month. Itโs a complex build but the lack of consistency is leading to chaos. I got a fix once I contacted the maintainer. Last week, PEastman got to the bottom of a complex vsite issue. The root cause was ultimately that we didnโt minimize before assigning velocities. It turns out that the initial velocity assignment has a term thatโs proportional to the max force, and in a starting geometry with large forces, that tem becomes huge and so do the initial velocities. Worked on ensuring vsite exclusions are correct. Added some test cases to cover tricky cases. So that should cap off OMM vsites, next will be GROMACS vsites.But for the time being Iโm writing a ton of tests. Over the past year, Iโve written tons of scattered regression tests. Iโm working on replacing those with the cases from Simonโs regression tests. ย
PB Fitting related work digging into aryl halogen training set differences between Gen1 and Gen2. QCA work, adding openff-default to the SPICE sets. DD reviewed over the weekend and these should be in queue Sitting on meetings with Relay folks, timemachine seems to be a faster alternative to openmm, possibly for fitting, something to test out sometime in future.
CC โ Did Relay folks have a specific shortcoming in OpenMM that motivated them to make timemachine? PB โ I think it was development friction - Contributing changes to OpenMM requires PEastmen review, and also often requires rewrited of lots of GPU code. Mentioned that they copied a lot of OMM code into TimeMachine. MT โ Itโs also that TimeMachine allows for analytical derivatives, which YTZ uses a lot.
Did some refits and passed a starting point for biopolymer fitting work, that includes improvements over Sage (fitting-exp/iter24 at main ยท MobleyLab/fitting-exp
CC PB passed me a starting point for biopolymer work (posted above). So Iโm working on getting ready to do protein fits starting from those. Iโll first fit the โnull modelโ , with the full QM dataset, but without any protein-specific terms. Then Iโll try some fits targeting some protein torsions. JW โ Do you already have librarycharges ready? CC โ Yes, Iโve got them using SMARTS from the ff14SB port, with the charge values from the biopolymer-ff call from a few weeks ago.
Continued to make progress on LiveCOMS review. Passing back to MGilson this week and expect this to move forward. ย ย
JW Worked a lot with JMitchell on toolkit showcase Tried and failed to help Hogan, review Horton PR Got icodes working somewhere around St Louis. Right angle torsion confusion
MT โ It would be good to identify the concrete issues here and who is handling each one. Also, each issue may require separate discussion as to whether is really is unintended. JW โ Will add to lead team agenda for this afternoon.
Ad board meeting - Was prepared for drama, was disappointed. We asked folks whether we should get a Sage 2.1.0 release out ASAP and they basically agreed โnoโ.
DD โ First deadline for F@H interface is next week. I donโt think weโre on track to meet this. How bad would it be to not have this done in time? CC โ We set this deadline in such a way to not be a blocker for my project. Iโm looking to have a preliminary FF ready on Oct 1. DD โ Ok, the ASAP deadline is also Oct 1. If I had to guess itโs about a 50% chance that weโll have this done for that deadline. CC โ The big deadline seems to be the NIH renewal in March. Itโs not clear to me that we need protein-ligand benchmarks done for that renewal. DD โ There are a few ways that we can have the necessary functionality ready, even if the entire stack isnโt fully featured. For example, the alchemical networks can work without the protien-ligand-benchmarks repo being at a final state. Iโll continue to check in with you to ensure that we have the pieces you need lined up. CC โ Sounds good, Iโll bring this up at the biopolymer FF call this week. I see two questions: When do we want to have any numbers at all coming out? And when do we want a production-ready repo available for anyone to pick up and run? JW โ I see this Oct deadline to be a technical deadline, where we get everything running in an automated way, and then weโll continue to monitor and bugfix based on what we see in production. DD โ Agree, thatโs my vision of this as well. The โtechnical-firstโ level of delivery is also good for ASAP, since they need to be running these calcs immediately, giving them a precise way to get relative numbers/rank order, and then they can focus on improving the absolute accuracy.
Thanks PB for putting the default-spec SPICE sets together. Looks like weโre through 180k calcs already.
PB + JW โ That impressive, the total size of all the spice sets is like 800k-1M calcs. DD + PB โ I notice that thereโs double the number of calcs here, which is due to the separate dispersion correction calcs. DD โ CC, Iโll reprioritize your jobs, I just submitting this at a higher priority to get a measure of runtime.
|