iiinova — where your digital self keeps evolving

Hannah Wayment-Steele @hkws.bsky.social · 21/08/2026

@ginaelnesr.bsky.social and I are excited to release makeshift -- a lightweight package to lower the barrier for using NMR dynamics data in ML. We hope this can pave the way for more NMR-based models like Dyna-1! Preprint: biorxiv.org/content/10.6... Docs: makeshift-docs.readthedocs.io

095

Hannah Wayment-Steele @hkws.bsky.social · 03/06/2026

Last weekend the W-S lab had our first lab retreat. Activities ranged from kayaking to vibe-coding science games! Thanks to Kemp station @uwmadisoncals.bsky.social for having us!

160

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

Big congrats to Yuchen Xu for leading the lab's first preprint!! And many thanks to @uwbiochem.bsky.social for their support! preprint: shorturl.at/ubA65 code: github.com/WaymentSteel... colab: shorturl.at/eSbQK

shorturl.at

141

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

Pratyush Tiwary’s group recently showed RNA LM embeddings are more informative than logits, our conclusions hold for their REDIAL method too, i.e.: 8/

110

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

One way we thought of to do this is to make a bunch of synthetic RNAs that fold to the same structure. When we did this for MEG3 RNA whose coevolution has been contested, we see it doesn’t have stronger CJ signal than the syn controls. But for preQ1, the CJ signal is stronger. 7/

100

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

We think gLMs have figured out how to store a lot of information, even to mimic thermodynamics. Even though it doesn't seem like “true” physics, it still poses a problem: we can’t take CJ signal as direct evidence for coev without more controls. 6/

130

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

Could it be that genomic language models “learned” the thermodynamics encoded in nearest-neighbor models? Our simple “mirror test” says no: nearest-neighbor physics models, if given a reversed structure, mirror their predicted structure. But gLMs don’t! 5/

110

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

We also saw CJ correlates with ensemble-averaged p(base-pairing) over all possible bp positions ?! 4/

100

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

We compared CJs from RNA-FM, Evo-2, and gLM2 to R-scape, a common method for RNA coev. We also compared to nearest-neighbor models for RNA ensemble properties. We saw CJ sort of matches R-scape, but agrees just as much with physics-based base-pairing with no coev support! 3/

110

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

In protein LMs, these signals are widely thought to reflect coevolution. We were curious if the same is true in RNA! Coevolution is a lot harder to detect in RNA and controversies exist over certain lncRNAs. 2/

100

Hannah Wayment-Steele @hkws.bsky.social · 01/06/2026

In the W-S lab's first preprint, we describe how genomic language models know something about RNA thermodynamics. Though we think this is cool, things get tricky! A growing practice for interpreting LMs is to perturb input tokens, often called "Categorical Jacobian": 👇

13314

Hannah Wayment-Steele @hkws.bsky.social · 17/05/2026

For any #nmr chemical shift aficionados, today I ported UCBshift(-X) from Teresa Head-Gordon's lab to work in modern Google Colab. Sharing in case its helpful! @nmr900.bsky.social colab.research.google.com/github/HWaym...

0103

Hannah Wayment-Steele @hkws.bsky.social · 15/09/2025

Cricket and Willow hosted a W-S lab dinner!

061

Reposted by Hannah Wayment-Steele

James Fraser @fraserlab.com · 20/05/2025

We are looking to hire (yes, even in this economy!) a jr. specialist to train in protein prep/structural biology related to our AVOID-ome work as part of openadmet.org. A great position for someone who is looking to be a tech for a few years before grad or med school. aprecruit.ucsf.edu/JPF05424

aprecruit.ucsf.edu

Junior/Assistant/Associate/Full Specialist Positions Available

University of California, San Francisco is hiring. Apply now!

02418

Reposted by Hannah Wayment-Steele

Stephanie Wankowicz @stephanieaw.bsky.social · 31/03/2025

Join us for the 2nd Macromolecular Conformational Ensembles Conference on June 9th/10th at UCSF. The most compelling questions in structural biology cannot be effectively addressed using only a single structure. conformationalensembles.github.io @fraserlab.bsky.social

conformationalensembles.github.io

Conformational Ensemble

22313

Reposted by Hannah Wayment-Steele

Gina El Nesr @ginaelnesr.bsky.social · 12/05/2025

someone said we need a meme

1296

Hannah Wayment-Steele @hkws.bsky.social · 09/05/2025

btw in the name of getting NMR dynamics data out there, RelaxDB is available at the Dyna-1 huggingface! Go check it out huggingface.co/gelnesr/Dyna-1

huggingface.co

gelnesr/Dyna-1 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

061

Hannah Wayment-Steele @hkws.bsky.social · 09/05/2025

Did some quick curation for ~30 of 133 proteins in our new dataset RelaxDB, which is the first of its kind to gather experimental info on timescales of motions per residue. The x-axis is intentionally silly to make a point - these dynamics expts do not use more than 10% D2O in samples.

140

Hannah Wayment-Steele @hkws.bsky.social · 09/05/2025

@dereklowe.bsky.social The paper you cite discusses difference when entire solvent is D2O. It could be that D2O partitions away from protein when it’s there at 5% since the paper’s main point is H2O has increased interactions with proteins. Curious if anyone’s investigated that!

1121

Hannah Wayment-Steele @hkws.bsky.social · 09/05/2025

this is why many protein nmr spectroscopists only use 5-10% D2O! I hope this doesn’t deter people from what imo is a super undervalued source of data

181

Hannah Wayment-Steele @hkws.bsky.social · 05/05/2025

👏 models in nmr structure != thermodynamics 👏 Structures I’ve been part of include the top-X most probable structures. Same as how the top-X-scoring models from a Rosetta run wouldn’t be a thermodynamic ensemble.

0140

Hannah Wayment-Steele @hkws.bsky.social · 28/04/2025

If we wanna make accurate boltzmann samplers, we gotta know what distributions they should be sampling 😁 These change in the presence/absence of ligands, and there are multiple systems for which this is well-understood by this point experimentally.

091

Hannah Wayment-Steele @hkws.bsky.social · 28/04/2025

BioEmu actually doesn't pass this test for AdK: majority of samples are in closed state (1AKE), which is the same intrinsic bias that AF2 + random sampling gets. I don't know the other apo/holo systems bioEmu looks at as well, but same story: sampled density is primarily at the holo state.

150

Hannah Wayment-Steele @hkws.bsky.social · 28/04/2025

You can see this in this FRET data for AdK here: in ligand-free form, the open state is more populated than closed.

100

Hannah Wayment-Steele @hkws.bsky.social · 28/04/2025

Really nice resource from @delalamo.xyz !! Wanna mention the "AdK test" we're noticing newer DL methods fall short on: many proteins (like AdK) that have apo/holo conf change sample both the apo/holo state even without ligand bound. But w/o ligand bound, they *are mainly in the apo state*

1144

Reposted by Hannah Wayment-Steele

Rita Strack @ritastrack.bsky.social · 10/04/2025

Your yearly reminder to acknowledge the core facilities you use and their staff scientists in your papers. These scientists are a crucial part of the scientific ecosystem and to continue to exist they need tangible credit for their work. Plus their associated expertise adds credibility to your work.

7484164

Hannah Wayment-Steele @hkws.bsky.social · 21/03/2025

I am sure future improvements exist over what we did! We removed deuterated samples (incomplete back-exchange) and entries with more than 12 15N assignments missing in a row. This is why we call this a "bold assumption", to me the proof is in the pudding, that we got a model with signal at all!

130

Hannah Wayment-Steele @hkws.bsky.social · 21/03/2025

Hi Gabe! Yeah we thought of af pair (no MSA) as upper limit but what you’re proposing would also be control for pair rep. To me, the kicker things we want to predict are coordinated motions that have low prob. Saw hints of that with CypA. I don’t think people would think of that as flexibility …

030

Reposted by Hannah Wayment-Steele

Kevin K. Yang 楊凱筌 @kevinkaichuang.bsky.social · 20/03/2025

Unassigned nitrogens in nmr data often indicate biologically relevant motion in proteins, and this can be used train deep learning models of protein dynamics! Hannah Wayment-Steele @ginaelnesr.bsky.social @sokrypton.org www.biorxiv.org/content/10.1...

1285

Reposted by Hannah Wayment-Steele

Gina El Nesr @ginaelnesr.bsky.social · 20/03/2025

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! 🧵

610538

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

Thank you so much to the amazing Doro Kern for dreaming big w me. Thank you to @ramith.fyi, Hasindu, and @sokrypton.org for pushing these ideas in early days! last but not least, thanks to @jcchildsfund.bsky.social and @hhmi.org for funding :))

141

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

Tremendous thank you to partner-in-crime @ginaelnesr.bsky.social. This collaboration started when she offered to clean a metadata spreadsheet, and she ended up pushing the deep learning we tested so much further than I could have alone!

140

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

Moral of the story: useful data is out there at many degrees of quality, but we need to know how to interpret the data. We're so excited to see where these models and data go next! Paper: rb.gy/de5axp Dyna-1 colab (thanks to @ginaelnesr.bsky.social ) and RelaxDB: github.com/WaymentSteel...

github.com

GitHub - WaymentSteeleLab/Dyna-1: Model for predicting micro-millisecond motions from protein sequence and/or structure

Model for predicting micro-millisecond motions from protein sequence and/or structure - WaymentSteeleLab/Dyna-1

150

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

Dyna-1 has predictive power in the big dogs of dynamics experiments: CPMG relaxation-dispersion. Dyna-1 predicted high p(exchange) in some aa's that typical data treatment says has no Rex, but more careful consideration says is real (NMR aficionados: unsuppressed R2).

130

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

There are so many interesting things that Dyna-1 predicts, but I wanna talk about a trend it didn't predict in RelaxDB! We realized that in many RelaxDB datasets where Dyna-1 did poorly, the Rex came from phosphate buffer binding/unbinding to the protein, an experimental artefact.

130

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

We split the mBMRB into train, val, and test set, and held out RelaxDB too as eval set. Sure enough, many pre-trained models that we tried - AF2, ESM2, ESM3 - had predictive power for Rex in RelaxDB. The best model we found was a middle layer of ESM3. We named it Dyna-1. Here's NtrC from test set:

140

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

We made the bold assumption that missing assignments were aa's with µs-ms motion. If we could train a model that could predict those, would it have learned µs-ms motion?

130

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

I next curated the "missing BMRB": 9,381 proteins with the aa labels: 0: 15N assn is present, 1: assn is missing. These labels are quite distinct from other properties like b-factors and missing residues in X-ray and EM (see paper for more).

130

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

Some aa's in RelaxDB didn't have Rex data bc they were missing from assignments, step 1 of any NMR study. This can happen if a peak is exchange-broadened due to µs-ms motion. AA's missing assns were also conserved. Then I realized - we have that data for >10,000 proteins, way more than 133!

140

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

I found a trend that amazed me in RelaxDB: aa's with µs-ms motion were more conserved than aa's with no motion, which were more conserved than aa's with ps-ns motion. **Depending on what timescale of motion we want to predict, we should expect different evol patterns.** What's the red category?

140

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

People have measured Rex for decades and these measurements have elucidated all sorts of phenomena, like allostery in this NtrC signaling domain. But there was no standard collection! Time to make a benchmark! Over 18 months, I curated 133 R1/R2/NOE datasets into "RelaxDB" (including NtrC).

140

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

Q: If I had an oracle for dynamics, what raw observable would I want to predict? A: Rex. Rex ⬆️ when atom's chem environment changes at µs-ms timescales, motions often critical for function. Rex ⬆️ if an alt state is present, even if we don't know what that alt state is.

150

Hannah Wayment-Steele @hkws.bsky.social · 20/03/2025

Protein dynamics was the first research to enchant me >10yrs ago, but I left in PhD bc I couldn't find big experimental data to evaluate models. Today w @ginaelnesr.bsky.social, I'm thrilled to share the big dynamics data I've been dreaming of, and the mdl we trained w them: Dyna-1. 📝: rb.gy/de5axp

28425

Hannah Wayment-Steele @hkws.bsky.social · 21/02/2025

I am so excited to talk about new science soon

030