2021-04-05 Polymer FF working session meeting notes

Recap

• Goal: Ultimately build a workflow that can take a bunch of monomers and the rules for how they can attach/grow from each other, and create a general FF that can handle any valid combination of those monomers.

• (General) – What’s the anticipated starting point for LW’s workflow? Is it a large parameterized chain? A definition of all the monomers?

• Can we fragment a large molecule to get the same benefits as knowing the underlying residues?

• What is gained by knowing the chemistry of monomers?

  • LW – It’s advantageous to know the underlying monomers to parameterize a variety of polymers. And visualization/analysis will rely on residue info.

  • JW – Knowledge of residue SMIRKS gives some guarantee of extensibility, so that sidechain amides won’t be confused with backbone

• SB – What’s the rationale for starting with a reverse-engineering approach?

  • JW – Largely a simple test case where we can get residue SMARTS early on, to be replaces with something more complex

  • SB – Could start with simple types, like smirnoff99Frosst, and build up complexity/elaborate SMIRKS as the data suggests

  • (General) – What’s the strategy for our own polymer FF? Not sure if we’re going to elaborate from small molecule FF, or simplify from ff14SB port.

Strategies for branching

• Scope of problem:

  • Types of branching:

    • Cysteine disulfide-ish = cyclizing

    • Dendritic polymers – See notebook from after last week’s second working session

      • 2021-03-31_working_session_2_cleanup.ipynb

    • Normal backbone addition

      • JW – It doesn’t seem meaningful to distinguish between “normal backbone addition” and “sidechain/dendritic branching”, since in the case of LW’s notebook, both possible attachment points are on the same atom.

•  

•  

Dealing with same residue in different environments (charges specifically)

• Two sorts of machinery that we’re building: enumeration machinery and ff creation machinery:

• Workflow

  • Monomer definitions

  • --(enumeration machinery)-->

  • a bunch of raw data (molecules+energies)

  • --(ff creation machinery)→

  • a force field

    • Note that it will probably be necessary for the “ff creation machiery” to decide that it doesn’t have enough information and call for another round of data generation

• Inputs required for running aspirational workflow:

  • monomer definitions

  • monomer attachment frequencies

  • backbone definition/attachment points

  • molecular context strategy (how far beyond a monomer a pattern should extend)

    • This should allow either a rigorous definition (“extend three bonds”) or looser definitions (“include the important context”)

      • Ex: “Residue N is very electronegative, and it affects the charges on atoms in neighboring residues. The FF creation machinery should somehow recognize this and have different parameter sets for monomers neighboring an instance of residue N”

    • Parameters in overlapping regions should be the same (dihedrals between residues shouldn’t disagree)

      • Ex: “If a polymer consists of residue A linked to residue B, and they both describe a proper torsion that spans their joining bond, then both parameters should be the same”

        • Given normal SMIRNOFF hierarchy rules, it might be a feature that one residue could override the junction parameter of the other

        • Alternatively, these special pairs that would have non-standard joining parameters could have backbone parameters defined using the entire residue structure of both neighbors

          • LW: This seems to me to be like the safe no-thought-necessary default

    • LW: IMO things that are difficult – extending molecular context much beyond the next residue/s. Difficult to account for 3D structure

• What are the pros and cons of having all patterns in the final FF contain at least one whole residue, versus just being the minimal chemical context to get the correct final parameterization?

  • It’s mostly bookkeeping – If we already somehow know which parameters will end up with physical values close to each other, then it makes no difference whether we’ve compressed them down into a minimal number of parameters/environments. But for a novel polymer, if we don’t know which substructures will be equivalent yet, it’s best to leave them separate during parameter training.

    • Machinery for parameter compression → ChemPer / SMARTY / Gokey’s current work

  •  

See if last week’s FF creation achieves the correct energies