Atlassian uses cookies to improve your browsing experience, perform analytics and research, and conduct advertising. Accept all cookies to indicate that you agree to our use of cookies on your device. Atlassian cookies and tracking notice, (opens new window)
Special thanks to David for arranging nearly all the logistics and shuttling us in his Sprinter van
Thanks to everyone who wrote a blog post. If you have any final edits to make, please get them in soon (or at least clearly communicate with me when you will do them) so I can hand all of these over to OMSF.
DD – Where can I see the results/notes from the meeting?
JW – Note that, since this update was recorded, the CCDC people got back to us and said that they’re not up to open source the datasets we’re interested in. So no CCDC datasets in the near future.
DC – Question about charge assignment - Does this avoid triggering AM1BCC charge assignment?
MT – If you ask to do AM1BCC on 5000 atoms, you will get a warning. I believe we set the cutoff at 100 heavy atoms.
JW – we have a function that assigns a dummy library charge to every smirks pattern, to make sure am1bcc never gets invoked at that stage
DD – You said there’s a tricky spot in the middle where you have to produce your own mapping of atom names. Is there a way around that in the future?
JW – Three answers:
1) If you have CONECT records, you can use the old style of PDB loading
2) What we show in the notebook, or some different useful approach that takes in a mapping from residue and atom names to cheminformatics structure in some form
3) It’s a little sketchy, but we are considering an approach that integrates with MDAnalysis or other guessers (guessing using coordinates, elements and total charge). We tried some tools that do this and weren’t satisfied with them, but we don’t like telling users no so we might make a way to allow users to do this deliberately.
CC – Any idea how many users come in with CONECT records vs. atom names?
JW – that’s an open question.
Rebranding NAGL
JE
NAGL is a GNN library we have used to train a model to reproduce AM1BCC charges
We have been referring to this model internally as “NAGL charges,” but the model needs a name distinct from the tool used to create it
Will probably need to name models as e.g. “nagl-am1bcc-feb2025”
Would like a catchy name to replace the “nagl-am1bcc” part, ideally one that could be adapted for future models such as e.g. “nagl-resp2”
Brainstorming ideas welcome!
MT – Point 3 is basically already the case for file names. I do use “sage” in conversation instead of “openff-2.2.1”, so I recognize the need to do that here.
DC – What does NAGL actually stand for?
JE – “Not Another Graph Library” I think
DC – To demonstrate the problem, consider something like NAGL$_{{\rm MBIS}}$(Q,$\mu$,V) - I’m writing a paper and am not sure how to name my model, and there aren’t naming/branding guidelines from OpenFF.
DD (chat) – Since we're using herbs for FFs, should we use another class of cooking ingredients?
JC (chat) – Citrus like, Clementine
MT (chat) – Peppers?
Shouts-outs
JE
Welcome Finlay Clark!
Positive feedback from the Interchange survey:
“Overall extremely useful tool, keep up the great work!”
“I have loved interchange and have introduced it to a bunch of friends who also are big fans now! Hopefully it's the future of MD.”
Q&A
Feedback on the watch party format?
JM – I like it.
CC – I would prefer to watch the video in advance
JW – maybe we should keep trying it out for a couple more months and then revisit the question
JM – If we are trying use social media to push these videos, we should stagger the release so we aren’t posting two or more videos on the same day. Space them out by a week
Also aim for under 10 minutes.
Keep them on a strict schedule, and post the schedule prominently
