Nate Corley

github.comGitHub - RosettaCommons/atomworks: A generalized computational framework for biomolecular modeling.A generalized computational framework for biomolecular modeling. - RosettaCommons/atomworks

1/n Crossover episode! We just released RFD4-Proteina, our latest protein design model - and the result of a year-long collaboration between the Baker Lab and NVIDIA GenAIR (the name says it all…). I hope you’ll agree - the results are astonishing. 🌐 Website: rfd4-proteina.org

10/n We also removed self-conditioning in favor of the now “in-vogue” flow-based reasoning via recurrence - a concurrent development! And we cleaned and scaled the data with AtomWorks (github.com/RosettaCommo...), of course

rfd4-proteina.orgrfd4-proteina.org

11/n Read the paper, explore the code, and let us know what you find: 📄 Paper: rfd4-proteina.org/assets/paper... 💻 Code: github.com/RosettaCommo... 🌐 Website: rfd4-proteina.org

8/n We designed tight and specific binders across a range of target difficulties, as well as two high-impact enzymes with direct comparisons to prior methods. And much more to come across other modalities in the coming weeks. Stay tuned!

9/n What makes it tick? We started with the Proteina-Complexa architecture and made a number of improvements. We first rethought triangle updates and improved how the single track communicates with the pair.

6/n Same story for enzymes. Even when including contemporary models like DeepMind’s recent AlphaProtein-novo.

7/n But does it work experimentally? For the applications tested so far, the answer is a resounding yes.

4/n Consider protein–protein binding, where we have the most baselines. RFD4-Proteina defines a new Pareto frontier.

5/n Small-molecule binding deserves its own highlight. Here, the in silico improvements are an order of magnitude.

2/n @kdidi.bsky.social and I started this project a year ago, then quickly joined forces with Max Kazman and Rafi Brent. A dream team. And what did we accomplish? A model that generates better designs at higher throughput than every baseline we tested. And by some margin.

3/n The in silico results shocked us at first. On some benchmarks, our clustered success rates are 3-5× those of the next-best model. And measured in successful designs per GPU-hour - like is often most relevant - it’s often an order of magnitude.

19 Sept 2025

RFdiffusion3 is here! We train a general network that explicitly models every atom and use it to design active enzymes and DNA binders. Tremendous team effort with Jasper, Rohith, Raktim, Rafi, Yanjing, Paul, Jonathan, and many others! Check it out: lnkd.in/eiUFfJaM.

Nate Corley reposted

Chaitanya K. Joshi

@chaitjo.bsky.social

RosettaFold 3 is here! 🧬🚀 AtomWorks (the foundational data pipeline powering it) is perhaps the really most exciting part of this release! Congratulations @simonmathis.bsky.social and team!!! ❤️ bioRxiv preprint: www.biorxiv.org/content/10.1...

Nate Corley reposted

Kieran Didi

@kdidi.bsky.social

Replying to @kdidi.bsky.social

`atomworks.ml` on the other hand offers advanced dataset featurization and sampling for deep learning workflows, all operating on the canonical AtomArray object from @biotite_python so that all transforms are traceable and generalizable between models. 4/6

Nate Corley reposted

Kieran Didi

@kdidi.bsky.social

Replying to @kdidi.bsky.social

AtomWorks has two main components: atomworks.io takes a file (cif, sdf, ...) and does parsing, cleaning and more. You can also look at your structures in a notebook or via PyMol thanks to pymol-remote, so you can directly inspect if your code does what you want! 3/6

Nate Corley reposted

Kieran Didi

@kdidi.bsky.social

AtomWorks is out! Building upon @biotite_python, we built a toolkit for all things biomolecules and trained RF3 with it. All open-source, test it via `pip install atomworks`! AtomWorks: github.com/RosettaCommo... RF3: github.com/RosettaCommo... Paper: tinyurl.com/y2w4z65b 1/6

biorxiv.orgAccelerating Biomolecular Modeling with AtomWorks and RF3Deep learning methods trained on protein structure databases have revolutionized biomolecular structure prediction, but developing and training new models remains a considerable challenge. To facilita...

(1/7) Training biomolecular foundation models shouldn't be so hard. And open-source structure prediction is important. So today we're releasing two software packages: AtomWorks and RosettaFold3 (RF3) [https://www.biorxiv.org/content/10.1101/2025.08.14.670328v2](www.biorxiv.org/content/10.1...)

(6/7) Frank DiMaio both directed and carried the team - he deserves the most recognition

biotite-python.orgBiotite documentation — Biotite

(7/7) We also want to give a shout-out to [Biotite](www.biotite-python.org/latest/), which is the bedrock of our approach. Biotite has made our framework vastly more performant and flexible. We're excited for what's next!

(4/7) There's no reason every researcher in the BioML space re-invents the wheel ever time they train a structure-based model. The complexity of loading and annotating biomolecular data for machine learning applications should be done once, and done right — that was our goal with AtomWorks

(5/7) It's been the pleasure my life to work alongside @simonmathis.bsky.social, @rkrishna3.bsky.social, @kinasekid.bsky.social , and so many other unbelievably talented individuals on this project. And extra credit to @kdidi.bsky.social for jumping into the frenzy to bring this work across the line

(2/7) We're thrilled to share RF3 fully open-source with the community — it handles chirality better than any other model, it supports arbitrary atomic templating (which we included at train-time), and it narrows the open-source gap to AF3