...
The second round of experiments is aimed at distinguishing between H and non-H atoms.
Both force fields have the following modifications (in addition to those described in iteration 1):
New parameter: a4a: [*;r3:1]~;@[*;r3:2]~;!@[#1:3] r3 atom - r3 atom - H
New parameter: a6a: [#1:1]-[*;r3:2]~;!@[#1:3] H - r3 atom - H
a13a: [*;r6:1]~;@[*;r5;x4:2]~;@[*;r5;x2:3] -->
[*;r6:1]~;@[*;r5;x4,*;r5;X4:2]~;@[*;r5;x2:3]
a42:[*;r4:1]-;@[*;r4:2]-;@[*;r4:3]--> [*;r4:1]-;@[*;r4x2:2]-;@[*;r4:3]
New parameter: a14a:[#1:1]~!@[*;X3;r5:2]~;@[*;r5:3]
Version 1 has expanded a8 and a9 to distinguish between nonring-ring-nonring vs ring-ring-nonring as well as distinguish more between H/nonH:
New parameter: a8a: [*;r4:1]@[*;r4:2]-;!@[!#1:3]
New parameter: a9a: [*;r4:1]@[*;r4:2]-;!@[#1:3]
New parameter: a9b: [#1:1]-[*;r4:2]-;!@[#1:3]
Version 2 has expanded a44 and a45 to distinguish between H/nonH:
New parameter: a44a: [#1:1]~[*;r4:2]~[#1:3]
New parameter: a45a: [*;r4:1]@[*;r4:2]~;!@[#1:3]
3-membered rings
First iteration of experiments
...
Which are both too specific (central atom must be C) and too broad (first atom could be in-ring or out of ring). Need to figure out how to work these together for the right coverage.
Looking at the distributions for a8 and a9 below, it’s not clear whether specifying ring-ring-nonring and nonring-ring-nonring separately will make a difference.
...
For now, I am trying two approaches:
...
Second round of experiments
...
First, I will change a42:[*;r4:1]-;@[*;r4:2]-;@[*;r4:3]--> [*;r4:1]-;@[*;r4x2:2]-;@[*;r4:3] to specify non-fused 4-membered rings.
Second, I will explore making both sets of exocyclic angle parameters more specific.
...
Add new parameters
a44a: [#1:1]~[*;r4:2]~[#1:3]anda45a: [*;r4:1]@[*;r4:2]~;!@[#1:3]to split out H vs non-H parameters
...
TODO: Might be worth putting 44 and 45 where 8 and 9 are in the order. Maybe they are picking up different things, conflating the comparison.
5-member rings
First iteration of experiments
Currently we don’t have any internal r5-r5-r5 ring angles, so I made one. I just made a generic one: [*;r5:1]@[*;r5:2]@[*;r5:3] but we may want to break it down further. Looking at the MSM parameter distribution, it seemed like the non-aromatic rings were clustered together, but the aromatic rings were all over the place in a way that made it not obvious how to split them.
New parameter: Added a new parameter to the end called a41.
Additionally, five-membered rings with S typically have a 90-degree angle around the S, rather than ~105 for other atoms. As a result I added a new parameter a43a with the pattern [*;r5:1]@[#16;r5:2]@[*;r5:3].
New parameter: Added a parameter a41a after a41.
I’ve also looked into splitting a13, as it currently covers both fused and spiro rings. Splitting them into two separate categories seems clear via the MSM parameters, so I added a13a ([*;r6:1]~;@[*;r5;x4:2]~;@[*;r5;x2:3]), which separates out the spiro rings. However, the split is less clear using Espaloma, as there is a lot of variation even within fused or spiro rings that is not present in the MSM data.
...
New parameter: Added a new parameter after a13 called a13a.Additionally, five-membered rings with S typically have a 90-degree angle around the S, rather than ~105 for other atoms. As a result I added a new parameter a43a with the pattern [*;r5:1]@[#16;r5:2]
Chris Bayly suggested looking into the ring-ring-nonring and nonring-ring-nonring parameters for 5-membered rings as well. I took a look and they didn’t look too different from the distributions they were a part of.
These external 5-member ring angles appear in almost every angle parameter distribution and usually aren’t very distinct. I think it would require a lot of care to separate them, as there is a lot of diversity currently being captured by the different parameters assigned to the angles, and I don’t want to lose that by lumping them together. Left for later.
Second iteration of experiments
Found a SMIRKs that captures all the 13a molecules: [*;r6:1]~;@[*;r5;x4,*;r5;X4:2]~;@[*;r5;x2:3]
...
a14 [*:1]~!@[*;X3;r5:2]~;@[*;r5:3] treats r5-r5-nonring--split into H vs nonH by introducing a14a:[#1:1]~!@[*;X3;r5:2]~;@[*;r5:3].New parameter: Added a parameter a41a after a41.
...
Issue with fused rings
One issue I have noticed with separating the small ring parameters is that there is no way to specify in a SMARTS pattern that a given atom is in a ring of a given size. The primitive r indicates the size of the smallest ring the atom is a part of, but if it is part of a fused or spiro ring, this may lead to issues. The primitive R denotes that an atom is part of a ring, but can only be modified by the number of ring bonds, not the size of the ring.
...
After a lot of experimenting I haven’t been able to find a solution that involves a single elegant SMARTS pattern. To get these right, we may have to add a number of very specific parameters, and increase coverage for fused rings.
Results
First iteration of experiments
Benchmarks for both versions of the Small ring FF are shown below. For DDE, Small ring v1 improves performance over Sage, and this improvement persists regardless of whether or not small rings are present in the benchmark set. This suggests that appropriately treating the small ring parameters leads to improvement in other parameters, that perhaps were pulled in an non-optimal direction to overcompensate for the incorrectly treated small rings. RMSD and TFD performance slightly improves over Sage, or stays the same.
I believe the worse performance of Small Ring v2 is due to grouping together H and non-H angles in a44 and a45, which are treated separately in Small Ring v1.
...
Other parameters I’ve looked at
...