Skip to content

Version 1.3.0 "Parsley"

Choose a tag to compare
@hyejang hyejang released this 21 Oct 17:49


This minor release contains a fix of amide-related issues; (1) a poor performance of v1.2 in reproducing amide torsional energy profiles and (2) absence of appropriate torsion parameters for dialkyl amides.

1. A poor performance of v1.2 in reproducing amide torsional energy profiles

Torsional energy profile of the amide bond of N-methyl acetamide generated using v1.2.0 failed to reproduce the QM torsional energy profile, forming a small hump at the QM minimum energy point (180 degree), which doesn’t appear in the QM profile.


Torsional energy profiles of N-methyl acetamide from v1.2.0

MM energy decoupling showed that torsion term contributed the most to the formation of the small peak, which indicated the need for the improvement of the amide torsion parameters.

From a closer inspection, we have found that the training set used in v1.2.0 fitting includes molecules whose geometries are not planar at a point where the amide bond is planar(dihedral angle = 0 or 180 degree) due to their large steric interactions or other chemical interactions, having a small peak at the point in their QM torsional energy profiles. This led the parameter set to be fitted to force the formation of a small peak at around their planar geometries even for simple molecules which don't have strong steric hindrance, like N-methyl acetamide.

So to avoid this problem, we carefully selected a training set for the amide torsion parameters with simple molecules which don’t have strong chemical interactions.

2. Absence of appropriate torsion parameters for dialkyl amides

Before the amide issue was raised, I found that there was a missing torsion term for dialkyl amides in the current parameter set. Due to the lack of appropriate torsion parameters for the dialkyl amides, it gives a low rotational barrier height, resulting in a bad reproduction of QM optimized geomeotries.


QM optimized geometry of Cc1ccc(=O)n(n1)CC(=O)N2CCCc3c2cccc3 (Magenta: MM optimized geometry with v1.2.0.)

To resolve this, we’ve decided to add new amide-specific torsion parameters(child parameters of t69 and t70). The preliminary study showed that the addition of the new torsion terms improved the performance of reproducing QM optimized geometries of dialkyl amides.

Here's the list of newly added amide-specific torsion parameters:

  • t69 [*:1]~[#7X3,#7X2-1:2]-!@[#6X3:3]~[*:4]
    • t69a [*:1]-[#7X3:2]-[#6X3$(*=[#8,#16,#7]):3]~[*:4]
  • t70 [#1:1]-[#7X3:2]-[#6X3:3]=[#8,#16,#7:4]
    • t70b [*:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#6,#1]
    • t70c [#1:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#6,#1]
    • t70d [*:1]-[#7X3:2]-!@[#6X3:3](=[#8,#16,#7:4])-[#7X3]

With this experience, we re-fitted the parameter set with the new amide torsion parameters to the carefully selected training set and confirmed it fixed the two issues.

Fitting Data and Method

1. Fitting Data

  • Torsiondrive target selection scheme: planar geometries (C-center improper dihedral angle, N-center improper dihedral angle between -5.0, 5.0 degree at the minimum energy point) + geometries which have local minima at 0 and 180 degree, selected from Gen 1 Roche torsiondrive dataset

    • # of selected targets: 62 1-D torsions
  • Optimized geometry, vibrational frequencies targets selection: molecules having t69s or/and t70s, selected from Gen 2 Optimization datasets

    • # of selected targets: 2347 optimized geometries, 532 vibrational frequencies targets

2. Method

Two stage fitting was carried out:
(1) fitting torsion parameters to the torsion training targets, then (2) reoptimizing bond and angle parameters to the optimized geometry, vibrational frequencies targets along with the torsion training targets used in the first stage.

First stage fitting was designed mainly to train the torsion parameters to the QM torsional energy profiles. The prior width of bond and angle parameters were intentionally set to 1/10 of the default scaling factor not to allow big perturbations during the optimization, and it used the default prior width of torsion parameters, so that it could be adjusted with more flexibility than other parameters.

The goal of the second stage fitting is to reoptimize the bond and angle parameters on top of the fitted torsion parameters in the first stage. The parameters were fittied to the QM optimized geometries, vibrational frequencies and torsional energy profiles, while keeping the torsional parameters from changing too much from the first stage by setting the prior width of torsional parameters 1/10 of the default width.

3. Fitting Result

In the first stage fitting, objective function decreased from 2.16e+02 to 6.29e+01 in 6 steps.

In the second stage fitting, objective function decreased from 1.72e+03 to 1.58e+03 in 14 steps.

Test Calculation Result

List of test sets:

  • Test set 1: N-methyl acetamide torsional energy profile;

  • Test set 2: All amide torsional energy profiles in the 2nd generation training dataset;

  • Test set 3: Optimized geometry, relative conformational energy benchmark full set;

  • Test set 4: Torsional energy profile benchmark primary set

The following table lists the single point calculation objective function values from different parameter sets for each test calculation, which give a rough measure of an overall quality of the performance.

test 1 test 2 test 3 test 4
v1.2.0 2.69 24.99 16681.14 384.88
v1.3.0 1.63 23.15 16436.44 404.51

1. Improvement in reproducing QM torsional energy profile of N-methyl acetamide


Torsional energy profiles of N-methyl acetamide

  • Clear improvement in reproducing the basis depth of QM energy surface near the minimum energy point (180 degree) over v1.2.0

2. Improvement in describing dialkyl amides


Torsional energy profiles of a dialkyl amide, C[N:3]([CH3:4])[C:2](=O)[C:1]1(CC1)c2ccccc2


QM optimized geometry of Cc1ccc(=O)n(n1)CC(=O)N2CCCc3c2cccc3 (Magenta: MM optimized geometry with v1.2.0. Green: MM optimized geometry with v1.3.0)

  • It also shows a better reproduction of rotational barriers and optimized geometries of dialkyl amides compared to v1.2.0.


v1.3.0 inherited the manual change made in v1.2.1 to resolve an issue with propyne substituents.