Loading, cleaning, and featurizing structural biology data for machine learning tasks is a HUGE pain. So we at the University of Washington's Institute for Protein Design developed a framework that makes it possible to prepare data pipelines for many bioinformatic applications in hours rather than months.

We then used this framework to train RosettaFold-3 - a state-of-the-art open-source model for biomolecular structure prediction.

Check it out!

AtomWorks GitHub: https://github.com/RosettaCommons/atomworks

RF3 GitHub: https://github.com/RosettaCommons/modelforge

bioRxiv Preprint: https://www.biorxiv.org/content/10.1101/2025.08.14.670328v1

Contributions welcome :)