2021-02-10 Meeting notes

Trevor Gokey

• TG: Tier 1 scoring: Slide 5
  I have a sp3-sp3 parameter I need o have a fine grained view of the gradient and I pulled the gradient for every IC. The gradient is technically zero at the minimum and yet contribution from each primitive would be non-zero, which I am looking at.
  CB: What does the dX^2/dp_i mean
  TG: These are all gradients with the decimals truncated. And this would give us an idea of what to split.
  B: these include all types of carbon carbon bonds?
  TG: These are the explicit terms and I have other decorators too.
  I iterate over all possible bits and partition my bits or in other words a split, and then check the difference in gradients, whichever gives the greatest abs. difference that indicates a better split
  DE = (2-5) = -3
  ABC = (1-1+3) = 3
  For this split score S1 = |-3-3| = 6
  DLM: You’re looking at the steepest gradient and that doesn’t mean that they want to go in that direction.
  CB: In your point 3, this is like a decision tree, this is like an old school optimization scan where there are too many correlated pairs, does this help distinguishing chemistry which includes more than one decorator at a time?
  TG: I talked about search depth before and I am showing only one layer here. Once I can join terms then I can look at contributing chemistries
  CB: Looking at the bullet point 2 you can qualitatively group A,C,D and B,E without going to point 3 and start looking at fingerprints. You already have information in point 2 and going to point 3 would increase the search space I think.
  TG: Yeah, you’re absolutely correct. The subtle point here is that there are certain bits that do not change the clustering and parameterization, and that’s when I have to add more bits and use a greedy approach, I cap that at 4. Looking at the negative and positive gradients and grouping may not be as clean, one example I can think of is a methyl group smarts, where one CH bond is clashing with other two protons.
  CCB: We are trying to avoid unnecessary searching. Moves in smarts space might waste time, can we do the search instead in chemical space. I see what you’re saying with the methyl groups. Did you end up finding nay non-optimized strcutures?
  TG: Fully optimized.
  CCB: You might be creating a smarts search so large and we didn’t even have an oxygen yet.
  CB: You’re right about the methyl group, but yet from point 2 you can look at separating E and C would always be the best outcome.
  CCB: There are tools that give you symmetry information, and you can get that information instead of searching every single bond then you would avoid these cases like methyl and butane having difference in proton behavior.
  CB/DLM: That’s a great idea.
  CCB: In essence instead of a single bond you can bin some of them and do a similar thing to gradient sums you’re doing.
  CB: I have worked on BCC and this didn’t occur to me before, great idea B.
  TG: I think the primitives would give the symmetry information you’re talking about
  CCB: Even if you have 30 heavy atom molecule instead of every single bond looking at the symmetry would help in grouping than smarts patterns.
  TG: I heard some great ideas for tier1 and I can try out other things too, one considering the largest difference in gradients instead of enumerating all grouped differences.
  CCB: We can come up with a dataset and do those experiments.
  TG: I can evaluate these at 5000 per second, so the real cost is in splitting and finding a split that puts the tier1 and tier3 scores on par.
  Thanks to CB, CCB, DLM for the ideas, will try out those.
  CB: You selected out a bond C-C, but in your previous proposal with butane you looked at all the bonds and this example makes it clear to me. Point 2 shows the various gradients negative and positive, large and small like the histogram PB was showing for his TIGs, this would give us an insight on what parameter to focus on. So, when we look at this single parameter our chemical space shrunk and we need a second tier of fingerprints that can help distill this subset.
  TG: That’s what I am going to do with the R-groups, and I can analyze different chemistries, writing up the bit-logic and working on the implementation.
  

• Tier 2 scoring: Slide 6
  Obj and gradients from tier1 scoring after split. I have a reference FF with n parameters before split and new FF after the split and look at the total gradient increase.
  CB: What if m >>> n, do you normalize?
  TG: I do not normalize and I am wrapping my head around it.
  CB: Keep a flag on this and when m >>> n and check whether this makes a difference.
  TG: Since we are making different splits the ref. stays constant and the highest first term would give us the highest S2 score.
  

• Tier 3 scoring: Slide 7
  This is the best score

• Initial conditions: Slide 8,9
  I am starting with a blank forcefield with just CH and CC, and the steps I discussed in my proposal on slack for my fitting procedure.
  CCB: In step2 are you using bond lengths and angles or splitting on k’s?
  TG: I need to have hessians to do that. I do geometry splits first and force constant fits later.
  CCB: Yeah, CB have mentioned something about hessians being sensitive
  TG: yeah, I took the hessians out
  CB: I think hessians are bad for this kind of work but that is the only source for force constants
  TG: The estimated k for bonds is only 6 units off, I looked with qubekit, it is 650 something.
  CB: Set the bond length to some number similar to perturbation and check.
  TG: I tried two variations and it didn’t work well.
  CB: Because of the torsion angle problem you couldn’t split on bond lenghts?
  TG: The drop was not enough, with torsions there is only 4% drop in objective and that means there is not much change for my parameters
  

• Fitting: slide 10, 11
  TG: Torsion splits are still not doing well, no new splits since the drop is around 4% and expected is 10%

• CB: Something is still suspicious to me, we may need to dig deeper, your objective function might be packing too much information and we may not be seeing the strongest contributions from torsions
  You might have made the bonds and angles really good and torsions are terrible from the get go and remain so.
  TG: Yeah, that’s what I am observing.
  CCB: Your optimized geometries are taking away the torsion information, may be adding torsiondrives might alter these results.
  CB: So when you are doing only bonds and angles did you get good results particularly with valence angles HCH CCH etc.
  TG: I think they’re doing good.
  CB: So it a success then
  TG: CCC bond split dropped obj. by 40% and gave a better FF.
  CCB/CB/DLM: Sounds promising.
  CB: Instead of brute force as in point 3 take the same problem and what would make it faster?
  What would it take to do clustering?
  TG: Gradients based splitting working well.
  CB: I think you proved that already.
  Use sklearn to cluster your fingerprints and do the split
  TG: It may not be a clean split if we are ignoring the bit vectors.
  CB: No we are not ignoring, we are using those feature vectors and chemper those to find the least common denominator. What if you have to do tier 3 relatively few times with the richer information.
  TG: Sounds good, I will work on those.