2021-09-03 WBO/Impropers meeting notes

@Pavan Behara

• PB: Initial steps on generating diverse set of conformers using MD.

• CB: There are better ways to sample different conformers, some doubts I have with MD are

  • whether it would invert the chirality?

  • in case of macrocycles even high temp MD is not good enough for those barriers

  • also, this doesn’t tell you if you fall short of sampling all minima

  • for example, with OE’s freeform you may sample much more minima and you are guaranteed that the chiral centers are inverted, also you must be careful with the default settings, try out Sage instead of MMFF94 to do a fair comparison with the MD results

• PB: Another question is whether we see conformations with energy differences of 50 kcal/mol? That’s what the MD traj is giving me.

• CB: My next question would be how do they look? You would find how they differ in 50 kcal/mol of energy from that. So, the goal is to actually see how many minima we come up with by doing MD or other?

• DM: The reason why we’re looking at the higher energy conformations is our fitting procedure includes only the minimum energy region and we are not necessarily penalizing the MM for sampling other minima, or for creating new minima that we don’t see in QM.

• CB: Ah, okay, this changes things then. There are certain kinds of high energy minima you need to look at, they’re not bonds and not angles, torsions and non-bonds. There is already lot of information in our laps with the torsion scans, where we gradually move up a steric interaction hill as well as non-bonded. But, do torsion scans include other non-equlibrium minima, that’re at off-minimum geometries, that’s a good question to look at.

• DM: I wonder if we do a constrained optimization on them by fixing all the torsions.

• CB: Yeah, PB what you’re doing with MD pick the same molecule and do a torsion constrained opt, which would leave you with some conformers, some of those might be redundant and can be filtered out by RMSD clustering, but essentially that would give a starting set of conformers that are away from minima.
  Take those 84 confs from omega-dense, constrain torsions, run MD with sage, then filter out redundant ones by basic clustering, then compare them with the MD ones. Then you would see which approach is faring well, and whether we need to run QM.

• DM: Yeah, sounds good. But, there is a chance these all fall near the same minima. I wonder whether this would still capture the problem of MM generating additional minima.
  I am still thinking of what would be a better criteria for success, it’s not like generating diverse set of conformers in a wide range of energy, the main purpose is whether we are creating spurious minima with our MM force field and how do we catch it/or avoid it.

• CB: That’s a great question DM, if we want to discover if our FF is giving out spurious minima, the best way is to use freeform dense sampling which does way more sampling than MD, and it finds minima analytically, and it would be a good way to find out the minima, and the next question is how many of these are spurious? Does it look like a good approach.

• DM: I think so. I am going to propose something like deliberately messing up the torsions by adding an extra term that is not supposed to be there, and whether the procedure finds any new minima.

• CB: That’s a great idea to distort the PES and sample the perturbed surface.

• DM: I was not proposing it like an algorithm to sample/generate weird minima, but to look at this problem of cases where we generate spurious minima.

• CB: I think that’d be good. I made freeform to tackle this and it is pretty fast. So, my idea is to do a restrained QM opt so that these do not fall into other minima, and then do a complete optimization and check whether they fall into the minima or remain on the walls of a PES. I am open to other approaches as well.

• SB: We can keep discussing these hypothetical scenarios, or may be we could go ahead and do some experiments and test instead of spending more time on it.

• CB: Another thing is, we are deliberately choosing highly stable minima so far in our training. So, strong sterics and spurious minima.

• DM: SO, the specific thing to try to look at is the number of minima coming out of the procedures, and whether we have artificial minima in-built in out FF.

@Jessica Maat (Deactivated)

DM: Welcome back from internship JM.

JM: Before my internship I was looking at the average improper angle Vs. the WBO for aniline dataset and there was this nice correlation between the two. SB was helping with 2D scans of this dataset, don’t know if it is complete now to perform a similar analysis or it would be complicated. I want to get feedback on what the next steps would be in this direction.

SB: We’ve got about 10 out of 24 2D scans complete.

DM: So, they’re hitting errors or ridiculously slow?

SB: Very slow.

DM: Do we have interpolated impropers yet in the toolkit?

SB: Don’t think so.

DM: Hmm, then it gets more challenging. How many developer hours do we have to spend on that?

SB: I am not exactly sure right now about that, may be we need to get JW and MT’s input on this.
One thing we could try though is using ANI to do these 2D-scans with a tighter convergence like 1e-10, I think Venkat from Cresset uses a pre-optimization with MM before doing ANI opts to get around convergence issues, so that might be a faster way to generate these 2D-scans.

DM: That’s a good idea. Since the toolkit implementation may take sometime we can look at cleaner datsets for wbo decoupled from impropers.

DM: As a second pass at WBO interpolated parameters we may do some work.

CB: Chaya’s work has a heartbeat for WBO interpolated parameters in biphenyls. We can look at the sets without steric interactions.

SB: My concern is we get these beautiful correlations by playing with the substituents at each end, what I would love to see is look at the industry benchmark set and whether we can see the same trend as well a continuous sampling of wbos with the impropers/or with proper torsions in the wild. If we don’t see many intermediate values and the sets are discrete then may be our approach might be different.

DM: Yeah, I agree that’s a good idea.

SB: I remember seeing tons of points around 1.0 or so with previous data.

PB: Yeah, we have stopped at this last time where we picked up a torsion specific to biaryls without ortho subs, and using SB’s constructure I tried to stretch the range of wbos this series can span.

CB: Yeah, we see other things overwhelming the wbo world, and in the real world all these are convoluted and we may never get out of it.

SB: I agree that inteprolated torsions add real value but I think more in terms of impact versus effort since everything is messy. Stuff that has more immediate impact on FF should be prioritized.

DM: Yeah, let’s take a look at the data we have and see the range of wbos it spans for this pattern and make a judgement call. JM/PB can work on this.
Second thing, to look at the non-interpolated impropers JM worked on before looking at the pyramidalization of Nitrogens. That’d be a good low hanging fruit.

CB: Also, along with the search for the wbos of biaryl non-ortho-sub smarts, look at the angle of those QM minima, and plot it. It’d be a surrogate measure.

DM: Yeah, it would be interesting to check that.

CB: Yeah, before we close, I completely agree with SB that we should focus more on things that have immediate impact on FF.