alisiafadini.bsky.social
Researcher. Interested in molecular biophysics using ML + protein structure experiments.
16 posts
229 followers
182 following
Conversation Starter
comment in response to
post
Hey Martin, yes — www.biorxiv.org/content/10.1... it’s a little (too far) down the thread!
comment in response to
post
It has been very much on my mind – was excited to see the code posted!
comment in response to
post
Very grateful for the work, support, and guidance of all authors: Airlie, Tom, @randyjread.bsky.social, @hekstralab.bsky.social, and @moalquraishi.bsky.social. It’s a privilege to work with such a great team. 14/14
comment in response to
post
Very interesting work is also happening using diffusion-based priors! 🔗Solving Inverse Problems in Protein Space Using Diffusion-Based Priors arxiv.org/abs/2406.04239 & Inverse problems with experiment-guided AlphaFold arxiv.org/abs/2502.09372 13/14
comment in response to
post
All method details and equations in our preprint: 🔗 www.biorxiv.org/content/10.1... Code available soon! 12/14
comment in response to
post
ROCKET performs a new type of structure refinement by optimizing latent representations in evolutionary space. This unlocks possibilities for high-throughput ligand screening, assemblies solved at low resolution, and conformational landscapes – automation 🔜 new frontiers. 11/14
comment in response to
post
Key Takeaways:
• No AF2 retraining needed
• Works with X-ray & cryo-EM/cryo-ET
• If you formulate a likelihood target, you can test with your favorite data type
• Refines large-scale conformational changes
• Robust at low resolution
• Enables automated experiment-guided refinement 10/14
comment in response to
post
ROCKET subsamples MSAs (inspired by www.nature.com/articles/s41...) to generate diverse starting models, selects the best-fit conformation, then refines further with gradient descent. Note: we cannot use pLDDT alone to succeed! 9/14
comment in response to
post
Gradient-based refinement struggles when AF2’s initial model is too far from experiment. E.g. AF2 predicts the serpin PAI-1 in a metastable active state, but experimental data shows it in a hyperstable “latent” state with a 40 Å loop shift. 8/14
comment in response to
post
The low-res challenge is key for emerging cryo-ET data.
🔹 ROCKET extracts two conformations from a 9.6Å GroEL map. Its modeling matches humans here and even surpasses them in tough regions, boosting fit to data from CC=0.2 to 0.5 in a flexible domain (see ⭐) 7/14
comment in response to
post
Model building below 3–4 Å is tough – even for experts.
🔹 ROCKET refines low-res (3.82 Å) HAI-1 X-ray data, improving backbone accuracy beyond AF2. It smartly preserves ambiguous regions & corrects a possibly misregistered helix (310–330), without adding geometric artifacts 6/14
comment in response to
post
✔️ Peptide flips (e.g. PTP-1B)
✔️ Domain shifts (e.g. GroEL)
5/14
comment in response to
post
ROCKET samples barrier-crossing conformations that standard refinement methods often fail to reach:
✔️ Ligand-induced loop rearrangements (e.g. c-Abl kinase and PTP1B)
4/14
comment in response to
post
ROCKET integrates experimental likelihood targets within OpenFold’s differentiable prediction pipeline to optimize MSA profile features. Structure refinement becomes a search in evolutionary space instead of Cartesian space. What does this unlock? 3/14
comment in response to
post
AF-based methods encode rich structural priors but lack a general mechanism for integrating arbitrary data modalities. ROCKET tackles this by optimizing latent representations to fit experimental data at inference time, without retraining! 2/14