PB: Even with a bad parameter set in the case of alkanes the objective function (X2) is lower than that with a final good parameter set, and the MM minimized structures overlap with the QM structures perfectly in both cases.
The average angle from QM for a1=110 deg., and a2=107 deg., and yet in case of bad parameters they drop down to 88.7 deg., and 88.1 deg., respectively (referring to column 2 in the table).
My hypothesis is that having dihedral_denom non-zero would bring parity in these two objective functions and there by there is a less chance of bad parameters having a lower objective function value and thus lower chance of being a global minima.
In case of sulfonamides, a similar situation where with good initial values (of Frosst99) and optgeo only targets, the X2 value is slightly higher at 671.66, whereas in another fit with bad initial and final parameters the X2 value is lower at 667.67 thus it makes sense for forcebalance to be happy with this minima.
CB: This was exactly the symptom that was plaguing charge fitting back in the 90’s and it is because of the near linear dependency in the fitting matrix. There might be other sources as well.
Also, there could be few contributions from sulfonamide targets and all others are contributing heavily, this is what I mean by near linear dependency and I used restraints back in those days but I know where to place them.
DLM: One thing that excites about modified seminario method is that we will get good initial values and may try to overcome multiple minima.
CB: PB, can you look at the relative energy differences and contributions to objective functions with sulfonamides only.
PB: These fits are with sulfonamide matching molecules only.
CB: So, you say that these molecules are moving these numbers to strange numbers, that seems fishy.
DLM: So, the puzzle here is that we are getting bad geometries and bad contributions in the objective function but they’re not showing up somehow?
SB: Yeah, that’s why PB was proposing to include dihedral rmsd because we may have perfect bonds, angles and impropers but bad dihedrals.
Also, from our previous experiments we did observe that the minimized geometries are fine and we are seeing the pathologies in MD.
PB: Yeah, when we looked at the CDK2 JACS series even with the bad parameters the minimized structures are sensible.
CB: What I remember from Hyesu’s presentation was that she observed pathologies with these parameters.
DLM: PB, flip through the geometries and their obj. function contributions and see if you can find all the angles are sensible and yet result in bad geometries due to dihedrals and things like that.
CB: Other thing I am interested in is that a30, a31 are quite generic and involve lot other chemistries, how they’re contributing?
