ROCKET came out on Nature Methods today. It takes a tremendous amount of effort to translate a research concept into a practical tool—one that researchers can seamlessly drop into their existing pipelines. We learned a lot along the path and will carry that spirit forward.
It was a fantastic trip and a pleasure talking with the Pfizer team 😄 One highlight was discovering they’ve built a web GUI for ROCKET to make it more accessible—I’m thrilled (and even a bit surprised) to see how our tool helping people solve real-world problems.
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
New TTPD with @fraserlab.com! We chat about travels, new preprints, and vibe coding.
podcasts.apple.com/us/podcast/b...
open.spotify.com/episode/6QCR...
We are releasing modular, clean APIs for these SW losses supporting both PyTorch and JAX. The motivation is to help you easily drop our implementation into your existing forward models and inference pipelines. 💻⚙️ (9/9) 🔗 t.co/qtNv87uBQj
While we tested this on Cryo-EM, the math is general. 📐
For inverse problems where signals don't overlap (e.g., medical imaging, astronomy, fluid dynamics), this differentiable, distributional framework can possibly help smooth your loss landscape too. (8/9)
We validated this approach in an end-to-end pipeline using Zernike3D.
On simulated datasets, the SW loss successfully inferred the heterogeneity landscape, producing a cleaner representation (intrinsic dimension ~1) compared to MSE. (7/9)
The solution? The landscape appears jointly convex close to the solution. 🔧
You can treat background contrast as a learnable parameter in the optimization loop. By jointly optimizing noise levels, we can eliminate the geometric bias and recover the correct structure. (6/9)
We want to be transparent about the challenges. OT is mass-preserving. ⚠️
If background noise (contrast) is miscalibrated, OT will physically "stretch" the structure to match the mass distribution. We demonstrate this geometric bias using a helical spiral toy model. (5/9)
Does the proxy work for complex inference?
We tested it by jointly recovering pose, CTF, AND conformation from a single noisy image starting far from the truth. 🎯
While MSE flatlined (blue), our SW proxy (orange) successfully guided all parameters to the ground truth. (4/9)
However, exact OT scales badly. We explored differentiable Sliced Wasserstein (SW).
Our key finding? A projected CDF-based L2 norm—akin to the Cramér-von Mises distance—is an efficient, fully differentiable proxy. It preserves the smooth landscape without quantile inversion cost. ⚡️ (3/9)
Why OT? Standard MSE relies on signal overlap. If model is far from the target, gradients vanish. 📉
OT measures "work" to move mass, creating a wide basin of attraction. In 3D volume alignment: OT (orange) consistently recovered poses from random starts where MSE (blue) failed. (2/9)
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)
ROCKET's webinar on cryo-EM+crystal data-guided protein structure prediction is now online 🚀: www.youtube.com/watch?v=_29C...
Thank you to the many of you who attended and stayed overtime for more Q&A, and especially to @sbgrid.bsky.social for hosting us last week!