0.
...
Torsion profile of amide bond of N-methylacetamide using v1.2.0 fails to fit the expected minimum (parabola shape), giving reproduce QM torsion profile, forming a small hump at the minimum. Re-fitting to the selected torsion targets which rotate an amide bond couldn’t remove the hump.
...
From a closer look at the targets used to train the amide bond parameters(t69s, t70s), I found that some targets behave not as expected at around its their planar geometry geometries (having a small peak, instead of parabolic shape near the minimum), which leads the parameter set fitted to be trained the targets to fail to reproduce the simple torsion profiles with parabolic shape near the minimum (like N-methyacetamide).
So one of the easy fixes fix of the problem will be is using simple targets which behaves behave exactly as we expect near the minimum (planar at 180 and small steric hindranceamide bond at the minimum) and here’s what I did:
1. filtering non-planar structure at minimum geometry
explanation of scheme2 and scheme3
...
scheme 2 (fb-fit2):
...
since the 2nd generation torsion training set is too small to filter, i pulled Roche torsion set (1st gen) and filtered with various scheme.
scheme 1. check improper dihedral angle of input.mol2 (which is generated using openeye toolkits) take targets whose C-center and N-center improper dihedral angles are both between -5.0, 5.0 degree;scheme 2. check improper dihedral angles at the minimum energy point of the torsion profile, take targets whose C-center and N-center improper dihedral angles are both between -5.0, 5.0 degree;
scheme 3. check QM data to see if the profile has its local minima at 0 or 180;
scheme 4. targets passed either scheme 2 or scheme 3
fb-fit2 vs fb-fit2-2
...
SMIRKS | initial guessfinal | fb-fit2 | fb-fit4 | |
|---|---|---|---|---|
t69a | [*:1]-[#7X3:2]-[#6X3$(*=[#8,#16,#7]):3]~[*:4] | 2.5 (1+cos(2x-180)) | 2.5 (1+cos(2x-180)) | 2.516923679597e+00(1+cos(2x-180)) |
t69b | [*:1]-[#7X3:2]-!@[#6X3$(*=[#8,#16,#7]):3]~[*:4] | 2.5 (1+cos(2x-180)) | 1.652160092455e+00 (1+cos(2x-180)) | 1.799860521668e+00(1+cos(2x-180)) |
t70 | [#1:1]-[#7X3:2]-[#6X3:3]=[#8,#16,#7:4] | 3.459249459574e+00 (1+cos(2x-180)) + 1.356955617521e+00 (1+cosx) | 3.459249459574e+00 (1+cos(2x-180)) + 1.356955617521e+00 (1+cosx) | 3.458787578663e+00(1+cos(2x-180)) + 1.357031068121e+00(1+cosx) |
t70a | [#1:1]-[#7X3:2]-!@[#6X3:3]=[#8,#16,#7:4] | 2.5 (1+cos(2x-180)) + 2.0 (1+cosx) | -2.081991778915e-01 (1+cos(2x-180)) + 1.400460484703e+00 (1+cosx) | -1.615997113282e+00(1+cos(2x-180)) + 1.289607014774e+00(1+cosx) |
t70b | [*:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#6,#1] | 2.5 (1+cos(2x-180)) | 4.091207931356e+00 (1+cos(2x-180)) | 4.079468007826e+00(1+cos(2x-180)) |
t70c | [#1:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#6,#1] | 2.5 (1+cos(2x-180)) + 2.0 (1+cosx) | 1.301446828323e+00 (1+cos(2x-180)) + 9.476830965983e-01 (1+cosx) | 1.020531593037e+00(1+cos(2x-180)) + 7.399229872960e-01(1+cosx) |
t70d | [*:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#7X3] | 2.5 (1+cos(2x-180)) | 1.271374777009e+00 (1+cos(2x-180)) | fb-fit2-21.142115082029e+00 (1+cos(2x-180)) |
...
Although the k1 value for t70a seems unphysical, the result below shows the plots got improved after the optimization. checked improvement for all of the four fitting targets.
...
SMIRKS | initial guessfinal | fb-fit2-2 | fb-fit4-2 | |
|---|---|---|---|---|
t69a | [*:1]-[#7X3:2]-[#6X3$(*=[#8,#16,#7]):3]~[*:4] | 2.5 (1+cos(2x-180)) | 1.679285484776e+00 (1+cos(2x-180)) | 1.778613695219e+00(1+cos(2x-180)) |
t70 | [#1:1]-[#7X3:2]-[#6X3:3]=[#8,#16,#7:4] | 3.459249459574e+00 (1+cos(2x-180)) + 1.356955617521e+00 (1+cosx) | 3.566161051804e-01 (1+cos(2x-180)) + 1.354248559088e+00 (1+cosx) | -1.357682774067e+00 (1+cos(2x-180)) + 1.266337423993e+00 (1+cosx) |
t70b | [*:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#6,#1] | 2.5 (1+cos(2x-180)) | 4.169159533591e+00 (1+cos(2x-180)) | 4.107144298045e+00(1+cos(2x-180)) |
t70c | [#1:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#6,#1] | 2.5 (1+cos(2x-180)) + 2.0 (1+cosx) | 1.240609067243e+00 (1+cos(2x-180)) + 9.065552775144e-01 (1+cosx) | 1.115929833406e+00(1+cos(2x-180)) + 7.646525300953e-01(1+cosx) |
t70d | [*:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#7X3] | 2.5 (1+cos(2x-180)) | 1.303139175986e+00 (1+cos(2x-180)) | 1.167960735598e+001+cos(2x-180)) |
...
For the test set (amide-rotating Roche torsion targets which are not selected either from scheme 2 or from scheme 3), can’t see the effect of splitting endo-cyclic rotation and exo-cyclic rotation
Parameter sets fitted to scheme 2 and scheme 4 performs similarly, while the parameter set fitted to the scheme 4 slightly more prefer to form hump at the near-planar structure. (could be good or bad)
fitting torsion parameters to simpler torsion targets (with less strong electrostatic interaction) might be needed for the next round fitting.