The MLSB workshop will be in San Diego, CA (co-located with NeurIPS) this year for its 6th edition in December 🧬🔬

Stay tuned @workshopmlsb.bsky.social as we share details about the stellar lineup of speakers, the official call for papers, and other announcements!🌟

is this how small molecules bind?? 😼

molecular cat-formations

Unfortunately, the MLSB Workshop @ NeurIPS (@workshopmlsb.bsky.social) was rejected this year.

Feedback from the deciding committee indicates it was a coin flip decision, with 283 proposals & a number related to “computational biology”

More on the future of MLSB soon…

Ran some sequence design with LigandMPNN and it... designed a protein with an (almost) His-tag 🤔

someone said we need a meme

10/ We evaluated Dyna-1 on RelaxDB.

Here are some examples comparing Dyna-1 predictions with NMR experimental data, including Olfactory marker protein, Protein Tyrosine-phosphatase A, and BPTI (and BPTI’s ms-timescale MD traj).

9/ The best performing model came from the intermediate layer of ESM-3: Dyna-1!

Notably, this model can take in a sequence, structure, or sequence & structure as input. While Dyna-1 performance is best using both, using just one as input still does well.

8/ Different pretrained models (AF2, ESM-2, ESM-3) all had predictive power. Inspired by several papers, I was curious if the embeddings from the intermediate layers of the language models had learned different relationships than the final layer

So, I did some sweeps…

5/ The mBMRB has ~10,000 proteins in it! What did we notice? For one, our dataset contained unique information compared to the PDB.

To start, 45% of the proteins had ≥1 NMR structure but no X-ray/EM structures

4/ But to predict dynamics, we need more data. Well, ms motion can cause peak broadening, so those aa's are missing assignments in the chemical shift data from NMR.

Where do we find these chemical shift datasets? The BMRB!

Hannah curated another labeled dataset: the "mBMRB"

3/ We also saw that AF2 pLDDT is lowest for aa's w/ ps-ns motion, and there was little discrepancy between aa's w/ no motion & µs-ms motion... interesting!

2/ From analyzing RelaxDB, we notice that there seem to be evolutionary trends depending on an aa's timescale of motion: aa's with µs-ms motions are more conserved than aa's with no motions (or ps-ns motions)

1/ We have large-scale exp observables for structures (PDB) and seqs (UniProt). So, what about dynamics?

That's Rex, a measurable from NMR. Rex increases when an atom moves at the µs-ms timescale.

There is no DB for that, so Hannah curated "RelaxDB"

Protein function often depends on protein dynamics. To design proteins that function like natural ones, how do we predict their dynamics?

@hkws.bsky.social and I are thrilled to share the first big, experimental datasets on protein dynamics and our new model: Dyna-1!

🧵

a good benchmark only cost me 22 CAD

thank you to everyone who came to @workshopmlsb.bsky.social!!! more so, thank you Erika Alden DeBenedictis for the fire crowd-sourced brainstorming 🔥🔥🔥

MLSB happening today! Packed room at @neuripsconf.bsky.social East Building Rm 11-12

@workshopmlsb.bsky.social