2021-08-27 Meeting notes

Participants

Goals

Figure out MD snapshots for conformational sampling

Discussion topics

Topic	Notes
Should we even do conformational sampling via MD	MD is about the worst possible way to sample conformers – CB It’s stochastically biased from the starting conformer 2 things make MD bad: 1. need to input energy to cross high barriers 2. Because it’s sampling from the Boltzmann distribution it’s going to get lots of low-energy conformers instead a diversity CB: Freeform: SMILES => conformers with (crude) partition function Can get results in the form as minima Can use these as inputs to ff parametrization Chooses many conformers, does minimization in (implicit? LW forgets) solvent, normal mode analysis to get minima DM: one thing I’ve done before is generate conformers, then start MD from each conformer, then cluster resulting conformers in implicit solvent Pro: if the conformer sampling gets near certain populations but doesn’t really get them, MD can characterize the region around them CB: hypothesis needs confirmation, especially in implicit solvent SB: porque no los dos? We can compare them. Makes the paper more interesting We don’t just want minimum structures CB: MD can deviate from the minimum well, if we start from FreeForm conformers SB: QM optimizations should be lightly restrained, e.g. as is done for ELF. Loosening tolerances can make this cheaper CB: Avoid thinking of it as starting with snapshots. Think of it as sampling diversity. You could deviate from minima with normal modes, that could be a good way We might end up with MD but should never start there
AIMD vs MMMD? Per-molecule vs general fit?	PB: Danny Cole says MM MD not useful. Not specifically sure how CB: always general, to avoid overfitting. Unless you want a very bespoke FF
What’s wrong with straight QM	DM: might wind up training to QM minima. MD might let you find other minima that exists in the FF CB: well you could just use a FF method w/o MD DM: What does omega do? How can we tweak things from the rotamer library? CB: I generate 40k conformers even if I only expect to use some hundreds CB: FreeForm definitely misses minima, but rarely important minima. We don’t need all minima for force fields, we just need a “decent” sampling of all minima we expect to contribute to the FF
Spitballing plans	SB: my preferred method: 1. FreeForm => conformers 2. MD to explore wells PB: we want sampling diversity. Can we try small displacements from conformers? SB: I think once in minima, do a short MD simulation 298K, then pick maximally diverse conformers CB: Danny Cole’s modified Seminario has a glorious destiny 🌅 Choose most diverse set (by some heuristic, e.g. torsions) Minimize in QM Using the modified Seminario to get 2nd derivatives Hessian immediately gives a pretty good snapshot of all internal degrees of freedom Gap: do we currently have a way to take Hessian and pull parameter info from it? SB: wants to systematically examine toggles of Hessian, internal coordinates, etc CB: perturbation from MD is hard bc valence parameters impose high barriers but we want the NB parameters Since these are local deviations, maybe a Hessian would do this well? SB: let’s write this up
Practical plans	DM: not enthusiastic about ML models at this time But could use it to help us pick snapshots SB: agree avoiding ML SB: ForceBalance’s model works well for ForceBalance but uses finite differences If we want to use autodiff then we need to backprop gradients through the minimization. This sounds unfun. Can we move away from some of the targets like evaluating at the minimization, and just do the QM snapshot SB: ForceBalance is fine for now ForceBalance questions: Should we include forces, cartesian vs internal coordinates, how should we contribute forces
What is the metric for success?	SB: benchmarking infrastructure for other valence refits Can we include properties like binding free energies RMSD etc metrics? CB: basic metric for success is alright QM minima, torsion fingerprint, liquid properties Caveat: special corner cases of those Specifically: I think we need to include steric interactions And highly repulsive electrostatics i.e. we need to cater for vdw and electrostatic corner cases Should we group distorted bond lengths, valence angles, etc., and require the FF to work well with those weird corner cases? DM: there’s asymmetry in how we optimize geometries with QM and get MM to agree with QM may have alternate minima in MM CB: is an example the weird sulfonamide geometries where we got “illegal” minima? DM: yeah probably is an example of this phenomenon DM: or, historically, in GAFF, the central aromatic ring buckles in the true minima. If you optimize it, it looks planar CB: omega would never generate the buckled ring, this is a disadvantage of Freeform DM: We should specifically look for cases where conformers in MD are lower in energy than the conformers that agree with QM minima. Basically looking for deficiencies in the MM landscape that you cannot find just by looking at QM minima
ELF10 vs not ELF10	CB: to validate ELF10 I have a collection of pathological molecules maybe got them from eMolecule finding ligands that will form strong internal polar interactions, e.g. h-bonds, salt bridges MMFF94 and AM1BCC comparator charge models took set of molecules and generated charges on each conformer of each molecule on each of the charge sets, found relative energies between all conformers looked at how relative energies internally differed between different charge sets used wanted to show if the internal energies were dramatically influenced by a single configuration with an internal polar interaction chose different ELF10 sets and showed relative energies of that were stable Will get set of molecules to Lily today (hopefully). Lily to ping on Monday if not CB: has given both posters and talks about ELF10 DM: could you please post on Zenodo for a DOI CB: would like to give a talk, maybe in one of these meetings CB: has given one at the Alchemistry conference