Fitting TIG* parameters

Simpler Fits

Simple_fit1

Replacing t43, t44, t45 with an interpolated torsion parameter for the smarts pattern: “[*:1]~[#6X3:2]~[#6X3:3]~[*:4]” and fitting this FF to 170 targets from 14 datasets (listed below under fit0) that have a dihedral that matches this pattern. The objective function value is compared to the zeroth iteration of various other FFs and fits:

Fit7

With slight changes to Fit4, TIG0 is converted to an interpolated parameter, and TIG1a and TIG1b are removed. Number of targets 928. Starting parameters are from simple_fit1 and fit4 values. Here are the parameters optimized:

['TIG0', 'TIG1c', 'TIG1d', 'TIG2', 'TIG3', 'TIG4', 'TIG5a', 'TIG5b', 'TIG6', 'TIG7', 'TIG8'] - interpolated

Fit4: Parameters optimized

• ['TIG0', 'TIG1a', 'TIG1b'] - General torsions

• ['TIG1c', 'TIG1d', 'TIG2', 'TIG3', 'TIG4', 'TIG5a', 'TIG5b', 'TIG6', 'TIG7', 'TIG8'] - interpolated

On the training set the objective function values are:

Fit4.1: For each of the interpolated parameter a general torsion parameter is created where the central bond can be a single, aromatic or double bond (denoted by letters p,q,r at the end of parameter id). Due to lack of enough training data that match those patterns only a subset of those are trained and here are the parameters optimized.

Parameters optimized

• ['TIG0', ‘TIG1a', ‘TIG1b',
  ‘TIG3p’, ‘TIG3r', ‘TIG4p', ‘TIG5ap’, ‘TIG5bp’, ‘TIG1cp’, ‘TIG6p’, ‘TIG7p’, ‘TIG8p’, ‘TIG2p’, 'TIG2r’, 'TIG1dp’ ] - General torsions

Fit 0

Input FF:

Open

Parameters to optimize:

• ['TIG0', 'TIG1a', 'TIG1b'] - General torsions

• ['TIG1c', 'TIG1d', 'TIG2', 'TIG3', 'TIG4', 'TIG5a', 'TIG5b', 'TIG6', 'TIG7', 'TIG8'] - interpolated

Targets:

1. 'Fragment Stability Benchmark'

2. 'OpenFF Gen 2 Torsion Set 1 Roche 2'

3. 'OpenFF Gen 2 Torsion Set 2 Coverage 2'

4. 'OpenFF Gen 2 Torsion Set 3 Pfizer Discrepancy 2'

5. 'OpenFF Gen 2 Torsion Set 4 eMolecules Discrepancy 2'

6. 'OpenFF Gen 2 Torsion Set 5 Bayer 2'

7. 'OpenFF Gen 2 Torsion Set 6 Supplemental 2'

8. 'OpenFF Group1 Torsions'

9. 'OpenFF Group1 Torsions 2'

10. 'OpenFF Group1 Torsions 3'

11. 'OpenFF Rowley Biaryl v1.0'

12. 'OpenFF Substituted Phenyl Set 1'

13. 'OpenFF-benchmark-ligand-fragments-v1.0'

14. 'SMIRNOFF Coverage Torsion Set 1'

Total number of targets excluding Lim Mobley benchmarks = 2746

QCA tdr_objects to exclude are in this file

Open

Fit 1

Input FF:

Open

Without excluding the in-ring torsions

Parameters to optimize:

• ['TIG0'] - General torsion

• ['TIG1c', 'TIG1d', 'TIG2', 'TIG3', 'TIG4', 'TIG5a', 'TIG5b', 'TIG6', 'TIG7', 'TIG8'] - interpolated

Targets: same as in Fit 0

Fit 2

Breaking up the interpolated parameters into single, aromatic and double (wherever possible) bond general torsion terms. Naming these as extensions of earlier TIG parameters appended by p, q, r for single, aromatic and double bonds repsectively. Wherever a carbonyl carbon is implied on the central bond there are no central double bonds, so not all parameters will have ‘r' extension. Excluding the high torsion barrier filters TIG1a, 1b so that double and aromatic bonds won’t get filtered.

Input FF:

Open

Parameters to optimize:

• ['TIG0', ‘TIG1cp', ‘TIG1cq', ‘TIG1dp', ‘TIG1dq', ‘TIG1dr', ‘TIG2p', ‘TIG2q’, ‘TIG2r’, ‘TIG3p’, ‘TIG3q’, ‘TIG3r’, ‘TIG4p’, ‘TIG4q’, ‘TIG4r’, ‘TIG5ap’, ‘TIG5aq’, ‘TIG5bp’, ‘TIG5bq’, ‘TIG5br’, ‘TIG6p’, ‘TIG6q’, ‘TIG6r’, ‘TIG7p’, ‘TIG7q’, ‘TIG7r’, ‘TIG8p’, 'TIG8q’ ]

Targets: same as in Fit 0

Fit 3

Corrected the phase of non-interpolated parameters (from Fit 2)

Open

Results of fits

Fit0 is better than 1.3.0 from the objective function values in the above table. Among CN and CC central bonds, CN has a lower objective function value and thus effect of CC is more dominant on the overall objective function.

Comparing MM Fits 0, 3 and 1.3.0 with QM

Fit 0 with all the TIG* parameters, and fit 3 is the non-interpolated version i.e., interpolated TIG params split into single, double and aromatic terms, compared with 1.3.0_unconstrained, and QM data.

Comparison is done on the training set of molecules, removing the ones with in-ring torsions and sorting the table based on the average of absolute difference in conformer energies between QM and MM_fit0. A full list of molecules sorted in ascending order of (QM - MM_fit0) can be seen at https://github.com/MobleyLab/wbointerpolation/blob/main/compare_forcefields.ipynb

Here is a list of top 5 molecules that are in very good agreement with the QM energies for the fit0 interpolated parameters FF:

	Torsion ID	Avg. abs(QM - MM_fit0) kcal/mol	Avg. abs(QM - MM_fit3) kcal/mol	Avg. abs(QM - MM_1.3.0) kcal/mol	Chemical Structure	QM-MM relative energies
491	{'tid': '1762178', 'assigned_params': {'fit0': 'TIG3', 'fit3': 'TIG3p', 'openff_unconstrained-1.3.0': 't47'}}	0.023835	0.405296	0.866851	Open	Open
6	{'tid': '21272427', 'assigned_params': {'fit0': 'TIG4', 'fit3': 'TIG4p', 'openff_unconstrained-1.3.0': 't43'}}	0.051240	0.397015	0.125842
76	{'tid': '21272438', 'assigned_params': {'fit0': 'TIG5b', 'fit3': 'TIG5bp', 'openff_unconstrained-1.3.0': 't43'}}	0.062916	0.274251	0.597345
628	{'tid': '21272422', 'assigned_params': {'fit0': 'TIG5b', 'fit3': 'TIG5bp', 'openff_unconstrained-1.3.0': 't43'}}	0.070763	9.416926	0.761913
626	{'tid': '21540566', 'assigned_params': {'fit0': 'TIG4', 'fit3': 'TIG4p', 'openff_unconstrained-1.3.0': 't43'}}	0.075898	0.410239	0.109622

Here is a list of last 5 molecules that have a higher difference in averaged MM energy with fit0 compared to QM: