2021-02-05 WBO/Impropers meeting notes
Participants
@Christopher Bayly
@Jessica Maat (Deactivated)
@Trevor Gokey
@Pavan Behara
@Simon Boothroyd
@David Mobley
Disclaimer: Missed noting some of the discussion on fingerprints
Discussion
DLM: Let’s continue the discussion on gradients from ff-release call
CB: To PB, the graphs look good but beyond that we can’t make any inferences without looking at the chemical structures of molecule populations that go in. There is a dire need to improve the visualization and presentation of results. I don’t like to look at just histograms and wait on further information. I want to see associated chemistries going forward in any graph you present, without that these discussions wouldn’t add any value. This work overlaps with what TG was doing with chemper.
DLM: Yeah, I agree. Create scripts where we can read in a sdf file and the SD tags with this gradient info, for example a plot with k1_bondorder1 and k1_bondorder2 on each axis and when we hover the plots we can look at chemical structures.
CB: Yeah, when you look at the plots related to TIG2 there are two peaks in k1_bondorder2 and only one peak in k1_bondorder1, I want to see whether the outliers in k1_bondorder1 would all correspond to a peak and things like that.
PB: Thanks for the feedback, will work on that.
SB: I can help with the visualization part.
CB: TG, checkout the clustering methods on scikit, something like dbscan would help. Also, coordinate with PB on getting his mols data.
TG: Okay, we will do that offline.
JM: I debugged the issue with the by-molecule fits when I tried to run on other datasets. I will add new data from these to my kval plots.
On other projects with the undergrads I have the results from the comparison of AM1 wbo from ambertools and off-tk, they look suspiciously good.
CB: Glad that you found the good results suspicious, Adrian Roitberg worked on ambertools implementation and compared his values with openeye AM1, he didn’t find any discrepancies and the results may very well overlap.
DLM: Yeah, but JM is doing an extensive study, so it would be good to double check.
JM: Okay, I will keep you posted if I find anything. Other thing I want feedback on is the comparison of conformer energies from XTB and AM1 calculations.
DLM/CB: Okay, you may try to use Victoria Lim’s scripts from benchmarkff where she compares the conformer energies with different forcefields, this is exactly the same thing except you’re using AM1 and XTB.
Another avenue to look at is WBOs from XTB and AM1, and how fast XTB is compared to AM1. Even though AM1 gives a crappy wavefunction BCC helps in correcting some of the charges. It would be good to know how XTB performs here.
JM: Okay, sounds good.