Access alternate functional conformations encoded in AlphaFold’s latent space in a few GPU minutes 👇 We also introduce a novel supervised transfer task: train once on a source (GPCR/kinase/transporter) and apply across the family. Work led by Minji!
The more challenging targets CASP gets, the more the structural biology community will learn!
Celebrating #BiophysicsWeek T&C partnered with Phase Space Invaders podcast: Robert Best @sonyahanson.bsky.social @xuhuihuangchem.bsky.social and host @milosz-wieczor.bsky.social Episode 32: Driving the Future of Biophysics: What’s Next in Theory and Computation?
www.buzzsprout.com/2313153/epis...
You shouldn’t have to have access to a GPU in-house to run ROCKET 💸 Phenix now hosts it on a server (free for academics rocket-9.gitbook.io/rocket-docs...). Try phenix.rocket and reach out if you need help – this is brand new! 15/15
You can now keep track of your ROCKET refinements in real time with WandB. Docs: rocket-9.gitbook.io/rocket-docs. Let one of our 1/🧑🚀s at @rs-station.bsky.social know if you run into issues at github.com/rs-station/... . 14/15
Grateful to the entire team, especially @minhuanli.bsky.social. Special thank you to @jovinelab.org, Suresh Banjara, Hiroki Okumura, Eve Napier, Pietro Fontana, Amir Khan for trusting us with their data and joining us in testing a new method. 13/15
Finally: ROCKET’s performance builds directly on @randyjread.bsky.social’s work on likelihood targets, experimental errors, and model bias. As we move into widespread data-guided prediction as a field, accounting for these effects is crucial for extracting signal and for model validation. 12/15
Which signals drive this stabilization? Disabling updates to high-impact residue pairs breaks recovery of the correct conformation – highlighting coevolutionary signals that underlie structural transitions in the model. 11/15
(iii) Experiment effectively informs prediction when structural context is missing. Example: this drug-induced loop transition is stabilized without predicting the ligand. 10/15
ROCKET outperforms human modeling, recovering the fg loop interaction with the neighboring chain – validated by follow-up higher-resolution data collection. 9/15
Real-world test 2: The ZPD egg-coat filament is poorly predicted by AF2. Modeling this 8.6 Å map, even with post-processing, is tough. 8/15
Real-world test: PPM1H-Rab8a partner protein interaction in the presence of severe preferred particle orientation. Relative domain placement validated through biochemical assays. 7/15
Extrapolated datasets capturing transient states can look much worse than nominal resolution suggests. Below, ROCKET recovers structural rearrangements of a time-resolved intermediate and models the bound state without explicit DNA. 6/15
We built controlled resolution benchmarks to test conformation recovery as signal degrades. One of them below: ROCKET remains robust down to ~10 Å. 5/15
ii) Signal-to-noise challenge. Modeling of high res data is largely solved. But as signal degrades (cryo-ET, time-resolved, flexible systems), recovering conformations is hard – even for humans. AF2’s prior helps complement low-res, noisy experiments👇4/15
i) Complex structural optimization is possible within AF2 embeddings, specifically at the level of coevolutionary features. In this space, barriers to structural rearrangements are reduced (see helix on the right). 3/15
ROCKET makes AlphaFold context-aware. We iteratively steer prediction to agree with experiment (cryo-EM, crystallography) at inference time, no retraining. Structure determination becomes a search where ML priors and experiment productively combine and inform each other. 2/15
ROCKET 🚀 inference-time optimization of AlphaFold to fit structural data is published! rdcu.be/fa9YH
Since our preprint, we’ve pushed it to regimes where other methods break: low resolution, weak signal, real experimental edge cases. Here’s what we learned: 1/15
New OpenFold3 preview out! (OF3p2)
It closes the gap to AlphaFold3 for most modalities.
Most critically, we're releasing everything, including training sets & configs, making OF3p2 the only current AF3-based model that is functionally trainable & reproducible from scratch🧵1/9
The Critical Assessment of Structure Prediction (CASP) experiment is calling for prediction targets: Immune Complexes, Organic Ligand-Protein Complexes, Nucleic Acids and Complexes, Conformational Ensembles, Difficult Protein Structures and Complexes. Rule of Thumb: If AlphaFold3 can generate a high-quality model, it is likely not a CASP-grade challenge. If it struggles, we want it.
Is #AI hitting a plateau in structure prediction? Help us find out at CASP17! 🧪🧬
Calling for Targets: Immune Complexes, protein - ligand complexes, RNA/DNA, conformational ensembles, membrane proteins, viral origins, and large complexes.
The Rule of Thumb: If AF3 can’t model it, we want it.
After an Amtrak coding session with @minhuanli.bsky.social, ROCKET team is back at @rs-station.bsky.social HQ! Thank you to Pfizer for welcoming us in Groton and to everyone who attended our seminar online. Exciting work ahead to continue exploring frontiers in structure determination. Back to it 🚀
Introducing The Structural History of Eukarya (SHE): The first proteome-scale phylogeny constructed entirely from 3D structure.
We computed 300 trillion alignments across 1,542 species to map the tree of life. 🧵👇 (1/5)
New Title Alert: meteor- a tool for computing crystallographic difference maps that specializes in robust identification of weak signals from minor populations such as bound ligands or time-resolved experimental changes.
Learn more here: buff.ly/bgJYF9N
#SBGrid #SBGridSoftware #StructuralBiology
Check our new preprint smoothing rugged Cryo-EM landscapes: shorturl.at/gYs9U
We tackle practical hurdles of Optimal Transport (OT) loss—differentiability, cost & noise sensitivity—make it a feasible inference workhorse.
W/ G. Woollard, D. Herreros, @pilarcossio.bsky.social, K. Dao Duc 🧵👇 (1/9)
The Inaugural Flatiron Institute Cryo-EM Conformational Heterogeneity Challenge pubmed.ncbi.nlm.nih.gov/41280101/ #cryoEM
A huge thanks to the team: TJ Lane, Virginia Apostolopoulou and Jasper van Thor, and to @rs-station.bsky.social for supporting the project! You can check out the full paper here: www.nature.com/articles/s42... or try it on your data: github.com/rs-station/m... 8/8
We hope this tool becomes a useful addition for the community working on dynamic systems, transient complexes, or small fragment screening. 7/8
Why? Many experiments (e.g. time-resolved crystallography 🎥+ ligand screening 💊) generate minor populations that are hard to find and model. Our method gives a way to detect them more reliably. 6/8
Xtallography can’t measure phases, though they carry key info. With TV denoising as a Bayesian prior, we infer these latent phases. Inspired by coherent diffractive imaging, we embed the TV denoiser in an iterative EM loop that fixes experimental amplitudes, updating phases. 5/8
