MSM as a better initial guess

Torsions and non-bonded parameters do influence the values of angles and bonds. Since we are performing a multi-objective optimization we will end up in a pareto-optimal solution where the parameter values might be nonphysical. Some of the earlier pathologies with bonds and angles that were fixed by hand are:

HMR issue with 1.2.0, fixed in 1.2.1

Open

It is to be noted that there was an error (or not) while building 1.0.0 where the force constants for both a single bond and triple bond were given the same value in SMIRNOFF99 FROSST, the initial force field port (force constants of 700 for both single bonds and triple bonds), and optimization didn’t help rectifying it down the line, which maybe due to priors penalizing large changes.

Sulfonamide issue with 1.3.0, rectified in 1.3.1

Open

Parsley parameters - From Hyesu’s summary (linked above)

Parameters a30, a31 are found to be going in the direction of nonphysical values over generations of force fields.

*equilibrium angle (degree) / force constant (kcal/mol)

“The equilibrium angles of a30 and a31 decreased while their force constants increased. And during the refitting of Parsley 1.2, the values moved further away from the very initial points (in SMIRNOFF99Frosst)

During the course of MD in a bound structure, NH2 of the sulfonamide is being pulled in close to the SO2 oxygens. The O-S-N angles go down to around 72 degrees and the distance between oxygen and hydrogen goes down to around 1.7 angstroms. Test simulation with a simple sulfonamide in water showed the same pathology, while it did not happen when using SMIRNOFF99Frosst.” - Hyesu

• Investigating the training set and test set sulfonamide optimized geometries didn’t reveal any pathology and they seem to agree well within accepted thresholds. This is significant since we cannot capture these silent errors unless we run MD.

• Also, in-ring geometries don’t have an effect even though they’re part of optgeo targets that target these parameters.

Nonphysical values for triple bond force constants

This is also closely related to the HMR issue outlined above and also how the values of force constants in SMIRNOFF99 FROSST and OpenFF-1.0.0 resulted in nonphysical but mathematically good values down the line.

MSM as a better initial guess

Danny Cole and Josh Horton’s initial work proposes to use modified seminario derived harmonic force constants for MM FFs. Josh built a 1.3.0 version of it using QM Hessians from Gen2 sets, and showed the differences between our current gen FF and MSM mean parameter values.

With this starting point using Mod-seminario angles and bonds brings parameters to physically intuitive values, and perform as well as Sage on the industry benchmark set. They also pass all the canary tests for HMR, which were placed as guard rails for any future releases to not relapse into it again.

***The parameter ids are still marked wrt parsley in the tables above

Optimized Force Field for this step

Open