2021-04-27 Bespoke Fitting meeting notes (sci)

10 mins

Fragmenter Update

@Simon Boothroyd

• SB: Fragmenter rewrite is done and ready for regression testing, the idea was not to reproduce the old version as this was not sustainable but we should have a method which produces sensible fragments.

• DM: I agree it does not need to exactly match the old performance but the fragments should be similar and sensible.

• JH: started running fragmentation regression tests between new and old oe versions also running ambertools version but this is much slower.

• SB: this should unblock bespokefit and qcsubmit

• DC: timing seems like an issue

• JH: this is due to doing sqm calcls on multiple conformers and rdkit generates more then oe

• SB: yeah we could do multiple sqm calcls at the same time which would help with speed and we could also use a NN to predict these WBOs in future. I also use ELF conformer pruning.

• SB: there is also no other sqm package we could use yet.

15 mins

ANI/geomeTRIC torsion scans

Venkat @Joshua Horton

Venkat: I have been looking into ani geometry convergence failures in qcarchive. I have extracted around 300 molecules that have issues from QCArchive removed molecules with missing element coverage. I then generate 24 input structures using rdkit and did a preopt using openmm then a full geometric optimisation using ani and default settings. I see around 93% success using this method.

Some molecules converge quickly some issue molecules take a a lot of steps to converge and visually they do not look good.

473 structures fail and belong to 156 molecules. Looking into the structures I see a pattern of where molecules get to a structure and the energy does not change much. The molecules are charged though and ani does not work well with charged molecules so these molecules need to be removed.

DC: we have been looking at trying to ease some geometry convergence issues, and they seem to be random sometimes where we get very close to convergence and have to randomly bump the coordinates to get us to converge.

DM: yeah this is a strange problem in the past I have looked for the atom with the largest force and randomly changed its coordinates.

DC: overall are you seeing a higher failure rate with ani vs QM

Vankat: I have not tried QM

DM: Was there something about ANI smoothness in the meeting earlier today?

JH: A group member from the ANI team presented earlier today on a new version of ANI2x that showed much smoother gradients.

DM: As I understand, it’s not released yet, right? We’re just going to help with testing.

JH: Correct.

DC: It may not be worth the effort to fix a problem that won’t be there in a month or two.

DC – JH, have you looked into bespokefit paper?

JH – Not yet

DM – There’s a process document. It specifies hard timelines for people to respond before each next stage.