1. Tracing the origin of the amide issue
...
One easy fix of the problem is using simple targets, whose torsion profiles have parabolic shape near the minimum (planar geometry at the minimum). + Expecting that this experiment to be one proof of the need for using simple-as-possible torsiondrive targets in general torsion parameter fitting.
2. Filtering out non-planar
...
structures at the minimum geometry
Several schemes to filter non-planar molecules:
...
(Since the 2nd generation torsion training set is too small to filter, i pulled Roche torsion set (1st generation) and filtered with various scheme. )
...
3. Comparison of different filtering schemes
...
Parameter sets fitted to Test set 1: (1) the simplest molecule, N-methyacetamide, (2) non-planar molecules filtered out in both 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)3;
Test set 2: all amide torsions in the 2nd generation training dataset;
Overall quality of fitting and test calculation:
v1.2.0 | fb-fit2 (scheme 2) | fb-fit3 (scheme3) | fb-fit4 (scheme 4) | Unfiltered | |
|---|---|---|---|---|---|
test cal 1 | 144.59 | 91.81 | 82.39 | 88.84 | - |
test cal 2 | 25.37 | 15.62 | 14.75 | 14.15 | 23.21 |
test cal 2-1 simple | 7.10 | 6.21 | 7.49 | 8.29 | 11.04 |
test cal 2-2 complex | 18.26 | 9.40 | 7.26 | 6.87 | 12.17 |
final X2 (fitting) | - | 1.31e+02 → 2.22e+01 | 1.30e+02 → 1.73e+01 | 2.09e+02 → 3.01e+01 | 5.28e+02 → 6.90e+01 |
(1) N-methylacetamide
...
(2) Test set 1 (N-methyacetamide, non-planar molecules in the 1st gen. Roche torsion set)
...
(3) Test set 2 (All amide torsions in the 2nd gen. training set)
...
(4) Check performance on non-planar molecules
...
(5) Direct comparison of final parameters
F(θ) = k1*(1+cos(2θ-180)) + k2*(1+cosθ)
ID, SMIRKS | Initial guess | fb-fit2 | fb-fit3 | fb-fit4 |
|---|---|---|---|---|
| k1 = 2.5 | k1= 1.6793e+0 | k1 = 1.3647e+0 | k1 = 1.7786e+0 |
| k1 = 3.4592e+0 k2 = 1.3570e+0 (from 1.2.0) | k1 = 3.5662e-1 k2 =1.3542e+0 | k1 = 2.4868e-1 K2 = 9.4280e-1 | k1 = -1.3577e+0 k2 = 1.2663e+0 |
| k1 = 2.5 | k1 = 4.1692e+0 | k1 = 4.4151e+0 | k1 = 4.1071e+0 |
| k1 = 2.5 k2 = 2.0 | k1 = 1.2406e+0 k2 = 9.0656e-1 | k1 = 1.5283e+0 k2 = 7.3406e-1 | k1 = 1.1159e+0 k2 = 7.6465e-1 |
| k1 = 2.5 | k1= 1.3031e+0 | k1 = 1.6149e+0 | k1 = 1.1680e+0 |
...
* About the counter-intuitive negative k1 value of t70, [#1:1]-[#7X3:2]-[#6X3:3]=[#8,#16,#7:4]
...
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.
...
* after fitting to planar molecules , does the force field force the non-planar molecules planar??
td_OpenFF_Group1_Torsions_071_C14H13NO- one example of molecules, which is non-planar at the minimum
...
4. Conclusions
Blah Blah
* Fitting torsion parameters to simpler torsion targets (with less + Selecting simple geometries (no strong electrostatic interaction) might be needed for the next round fittingmay need to be considered in the next torsion training dataset selection procedure.