2021-02-26 WBO/Impropers meeting notes

Owned by Pavan Behara
Last updated: Feb 15, 2022 by Diego Nolasco (Deactivated)
3 min read

Participants

@Jessica Maat (Deactivated)

@Christopher Bayly

@Trevor Gokey

@Simon Boothroyd

@David Mobley

@Pavan Behara

Discussion

JM: Here are the plots with tig proof of concept fit and almost all of them show a good linear trend wrt the matched molecules

SB: This looks good, what’s the difference between PB’s earlier fits and these?

PB: These are very specific parameters and JM worked on pinning exact chemistries, the patterns I used are too general to begin with.

DLM/CB/SB: May be we should benchmark this new fit and see how it fares on molecules other than substituted phenyl, PB can take up the task.

 

JM: I also have these ddE plots for comparing XTB and AM1 calculations, and XTB is superior here.

SB: That’s amazing, do you have an idea about computation time?

JM: I will check and get back to you.

 

JM: I want to show one of the high barrier molecules on the plot of my TIGs, it looks similar to the outliers we saw last week in the wbo toolkit backend differences.

SB/DLM: Also, Chaya’s earlier plots used ELF10 calculation for wbo and your current plot don’t use it yet, may be for consistency we can redo earlier plots with our current tk version wbos.

 

PB: For t96-97 interpolated parameters here are some splits, and from training on two molecules it seems having a C-halogen terminated smarts would be useful in some of these cases, they match the phase better than 1.3.0

DLM/CB: This looks okay, may be we can restrict the halogens to Cl, F and remove Br,I since they’re not as electronegative as F, Cl and almost behave like carbon in electron pulling capacity. Agree that we need more data points to make an inference.

PB: I started looking into clustering based on OE fingerprints within a single TIG parameter and here are some preliminary binnings of molecules.

CB: Look at the molecules and identify which ones are more electron withdrawing and which ones aren’t and that would guide you to make a decision about splits. TIG6 is too general and amides must be treated separately. Anyl groups here stand out separately, and there is strong conjugation when we have a bridge Nitrogen. And in some of these molecules there is strong steric interaction rising from ortho substituents or adjacent rings.

This approach of starting with too-general parameters is proving to be tedious, on the other extreme we have a very good fit with JM’s highly specific interpolated parameters, we need to strike a balance between these two realms. We can try to make some of the specific parameters more general and stop before it goes off the rails.

Few initial tasks would be

  • to check how well JM’s parameters do outside substituted phenyl set, which chemistries still hold good and which are going bad

  • check which other chemistries outside the substituted phenyl set of molecules are matched to each torsion parameter, and whether they bring the linear trend down or holds it good

  • another way would be to restrict PB’s TIG6 to match only JM’s set of chemistries/molecules

These exercises would help us determine how efficient our wbo method is. Also, remove strong sterics from the analysis so that we can remove the role of vanderwaals and electrostatics interactions from our analysis. I am not yet sure optimizing a global scalefactor would be beneficial but it’s worth a try.