2025-01-22 ZAFF meeting notes

Recording will be uploaded

JCl (slide 7): noticed some consistent relative differences between the two methods, but there seem to be two clusters with the relative differences consistent within. Did you notice any systematic differences between the clusters?

• JG: Thought that might have to do with halides in the ligands. Howver, 14 and 15 (outliers in UAFF) don’t have them. I haven’t looked into this further. What we concluded is that it would be nice to try switching up the small molecule FF.

JW: I’m curious how often do the bonds between proteins and metals exchange? It seems for ligands that many of them could be considered bonded, but in RBFEs you can’t model bonds. Can you differentiate between proteins where bonds exchange quickly and where bonds are fixed?

• JG: I think so. I haven’t experienced many other metalloproteins, but this may be individual and specific to each research problem. In our case, we know we should have tetrahedral Zn from experiment and don’t want to see any change in coordination.

LW: could you go into issues with waters coordinating to Zn in MCPB?

• JG:

• JG – Was a long time ago. In step 2, I think there was some sort of bug, and I talked briefly with AmberTools folks but not sure if the fix got into more recent versions. Working with someone now whos trying to use it so I may get to revisit this.

• LW – Any cases where you did end up getting it to work?

• JG – The ones where I ran into issues were nonbonded models. After that, we started thinking about making our own nonbonded model. So a combination of usability and how to run the models.

TM – I thought we got it to work for some systems?

• JG – Got it to work for the bonded model. But had issues with nonbonded options, might be a bug which has since been fixed.

JW: In the cationic dummy atom model, do the R-groups need to know they’re coordinating to Zn?

• JG: in this model you don’t change anything about the coordinating residues, you just apply the protein FF. In UAFF, you do have to change the name of the residues to get the correct parameters.

• JW: And the dummy atoms just have normal L-J and q, they don’t have any special polarisation?

• JG: yes

DC (chat) – Nice work! Have you thought about how much data would be needed to train a really minimalist ML FF? Eg something just modelling metal-organic atoms within a 5Ang cutoff

• TM (chat) – Yes, I’ve worked this out for an unfunded grant proposal. I can look it up.

• JG – Yeah, we’ve been talking about this. TM knows more about ML side.

• TM – Don’t have much more to say, grants on this topic haven’t been funded. I proposed looking at minimalist systems like MOFs or all zinc sites in PDB.

• JG – There’s a paper I’ve been looking at (Merz) where they did a SchNet model with EwaldMP method to account for long range interactions. Wasn’t for proteins, was for metal coordination complexes. Had 500ish metal coordination complexes from CSD. Was thinking about using a similar methodology but applied to proteins. But paper didn’t mention using sulfur coordination, which is a need for me.

  • https://pubs.acs.org/doi/10.1021/acs.jcim.4c00095 ?

• DC – Looking at some of these systems, when you look at all the ways you could apply constraints, it does seem like learning+building an ML framework would be simpler.

• TM – For something like a ZAFF style model, probably not too much.