Skip to content

Description

This release contains results for fitting with screened torsion profiles, based on hydrogen bonds. The result is highly based on work from @hyejang.

In the current version of the force field, the charge parameters are determined by AM1BCC calculations for each molecule. Since it is not guaranteed that these charge parameters can correctly reproduce strong electrostatic interactions, especially hydrogen bonds, we determined that the torsion profiles with hydrogen bonds should not be included in the fitting.

To detect the hydrogen bonds, we tried two approaches. First is to use the energies of electrostatic interactions. We tried to look at the absolute energies and relative energies along the torsion profile, and we found that this approach is able to find out most of the hydrogen-bonding systems, however it may also have false positive detections on peptides and other molecules with larger charges. Then we tried the geometry-based method, namely Baker-Hubbard (𝜃>120 and 𝑟<2.5A) and found that it gives a good result. Therefore we went for the second approach. (mdtraj package is used for the baker_hubbard function)

For implementation details, see PR #2. @hyejang has also put together a detailed document on the investigations about detecting hydrogen bonds, which is attached in this release.

Fitting Targets

This release contains the same fitting targets for optgeo and vibfreq targets as previous. The torsion profile targets are re-generated.

Roche set: 669 1-D torsion profiles; 936 optimized geometries; 660 vibrational frequencies;
Coverage set: 417 1-D torsion profiles; 831 optimized geometries; 235 vibrational frequencies;

The number of torsion profiles for Roche set decreased from 798 to 669, indicating that about 130 torsion profiles are filtered out because of hydrogen bonds.
The number of torsion profiles for the "Coverage Set" increased from 378 to 417. This is because some tasks previously in an "error" state got fixed on the server.

The parameter coverage is 481/500.

Results

The objective function decreased from 25708 to 4522. The convergence behavior is similar to previous releases, and the relative amount of decrease is a little higher than previous release, indicating that the screened torsion profiles overall are fitted better.

The fitting results for the bond and angle parameters are generally consistent with previous releases. The torsion ks have some differences, such as k1 of t127 and t155 now become negative.

Notes

In this release, the scripts are updated to be compatible with the new qcportal 0.9 and qcelemental 0.6.1.

Assets 4
You can’t perform that action at this time.