MC states not mixing in 0.16
See original GitHub issueFor nearly the same experiment (only difference is slight change in clearance
and existence of switch_distance
), MC states seem to not mix in 0.16
, while it seems to work in 0.15
(see attached YAML files)–i doubt these differences contribute to the overall result though.
yank_mc.zip
During yank analyze
in 0.15
, the reported state mixing matrix is (relatively) populated (see attached JQ1_15_analyze.out
), while in 0.16
, it is completely empty (see attached JQ1_16_analyze.out
). Furthermore, the following is also reported in 0.16
:
2017-07-17 07:59:50,232: Perron eigenvalue is 0.00000; state equilibration timescale is ~ 1.0 iterations
Unsurprisingly, the dG estimate is very, very wrong in 0.16
(~ +90 kcal/mol), and relatively correct in 0.15
(~ -15 kcal/mol). Interestingly, yank analyze
takesmuch longer in 0.16
(currently ~5 min vs. 20 seconds).
Issue Analytics
- State:
- Created 6 years ago
- Comments:41 (33 by maintainers)
Top GitHub Comments
Hydration free energies in TIP4P-EW seem to be working with the latest openmmtools now!
Thanks, @spadavec!
A quick manual inspection of
system.xml
suggests the cause: There are no virtual site definitions! The virtual sites have point masses set to zero, but instead of being updated by construction, they’re probably fixed in space because of the missing<VirtualSize>
block. I’m surprised the binding free energies weren’t even crazier!I think we neglected to properly handle virtual sites in
AbsoluteAlchemicalFactory
. This is one of the dangers of creating a newSystem
from scratch instead of using a deep copy ofSystem
as a starting point and then removingForce
objects (which is now possible)—a scheme I think we should switch our design to later on.@andrrizzi : We’ll need to add virtual site support.