Accelerating Biomolecular Modeling with AtomWorks and RF3

Nathaniel Corley, Simon V. Mathis, Rohith Krishna, Magnus S. Bauer, Terry Thompson, Woody Ahern, Maxwell Kazman, Rafael I. Brent, Kieran Didi, Andrew Kubaney, Lilian McHugh, Arnav Nagle, Andrew Favor, Meghana Kshirsagar, Pascal Sturmfels, Yanjing Li, J. Laurence Butcher, Qiang Bo, Lars L. Schaaf, Raktim Mitra, Katelyn V. Campbell, Odin Zhang, Roni Weissman, Ian R. Humphreys, Qian Cong, Jonathan Funk, Shreyash Sonthalia, Píetro Lió, David Baker, Frank DiMaiobioRxiv 2025 · DOI

TL;DR

This paper introduces AtomWorks, a flexible and robust framework designed to accelerate the development of biomolecular machine learning models. Using AtomWorks, the authors trained RosettaFold-3 (RF3), an open-source protein structure prediction model that significantly closes the performance gap with state-of-the-art closed-source models like AlphaFold3, particularly in handling complex biomolecular interactions and chirality.

What Problem Does This Paper Address?

Developing and training cutting-edge biomolecular machine learning models, especially for structure prediction and design, is a significant challenge. Existing open-source efforts are hampered by:

  • Data Heterogeneity and Engineering Overhead: The Protein Data Bank (PDB) and other sources contain complex, often messy data with numerous edge cases (e.g., misannotated bonds, multiple occupancies, ligands). Preparing this data for model training requires extensive domain expertise and custom pipelines, consuming months of effort.
  • Lack of Generalizability: Bespoke data processing pipelines for each model prevent easy reuse and lead to duplicated effort across the research community.

The authors aim to democratize biomolecular modeling by providing a unified, high-performance framework that simplifies data handling, promotes code reuse, and enables rapid prototyping of new models.

Methods

The core contributions are the AtomWorks framework and the RosettaFold-3 (RF3) model trained within it.

  • AtomWorks Framework:
    • Modular Data Processing: AtomWorks standardizes input processing, handling numerous PDB edge cases (e.g., removing leaving groups, correcting charges, parsing geometries) to produce higher-quality derived features.
  • Unified Representation: It uses an "AtomArray" (from the Biotite library) as a common, unambiguous representation for biomolecular structures, allowing for easy chaining of data processing operations (Transforms).
  • Research Velocity: This modularity and shared representation significantly reduce code duplication and enable researchers to focus on specific hypotheses rather than data pipeline engineering.
  • Scalable Training: The framework is designed for efficient and scalable training of various biomolecular models.
  • Industry-Grade Practices: It emphasizes comprehensive documentation and extensive testing (>85% coverage).
  • RosettaFold-3 (RF3) Model:
    • Structure Prediction: RF3 is an all-atom biomolecular structure prediction network.
  • Improved Chirality Handling: It explicitly models stereochemistry using dihedral angles and employs data augmentation (inverting chirality in 2% of examples) to achieve superior accuracy in predicting chiral centers, outperforming previous open-source models and approaching AF3.
  • Arbitrary Atom-Level Conditioning: RF3 allows users to specify constraints (e.g., distances between atoms) for templating, protein-ligand docking, or folding around specific ligand conformers.
  • Novel Distillation Datasets: RF3 was trained on new datasets, including RNA and protein-nucleic acid complex distillations, and a disordered region distillation set generated using Rosetta.

Datasets

  • Training Datasets:
    • Protein Data Bank (PDB) (chains/interfaces)
    • AF2 Monomer Distillation
    • Multidomain Distillation
    • RNA Monomer Distillation
    • PDB (Disordered regions)
    • Nucleic Acid Complex Distillation
    • Inverse Chirality (used as a data augmentation)
  • Evaluation Datasets:
    • Recent PDB Evaluation Set (structures released after the latest training cutoff, de-leaked against training data)
    • Antibody Evaluation Set (using SAbDab database and ANARCI for numbering)
    • Mixed Chirality L/D Macrocyclic Peptide Evaluation Set (from Cell 2022 paper)

Key Insights and Ideas

  • Data Quality is Paramount: The paper strongly emphasizes that robust, clean data processing is a critical bottleneck in biomolecular ML. AtomWorks's meticulous handling of PDB edge cases directly translates to higher-quality downstream features and improved model performance.
  • Modularity Accelerates Discovery: By decoupling data processing from model-specific logic, AtomWorks allows for rapid experimentation and reuse of components, significantly speeding up the development cycle for new biomolecular models.
  • Chirality is a Key Challenge (and Opportunity): Traditional methods struggle with accurate chirality prediction. RF3's geometric approach to chirality, combined with data augmentation, demonstrates that this can be effectively learned, leading to more accurate predictions for small molecules and mixed L/D peptides.
  • Open-Source Empowerment: The release of both the AtomWorks framework and the RF3 model aims to democratize access to state-of-the-art biomolecular modeling tools, fostering further innovation in the open-source community.

Key Results

  • RF3 Performance: RF3 demonstrates competitive performance with AlphaFold3 (AF3) on various structure prediction tasks, significantly narrowing the gap between closed-source and open-source models.
  • Chirality Prediction: RF3 predicts correct ligand chirality for 88% of centers without inference-time guidance, outperforming AF3 (84%) and Boltz-2 (76%). For mixed L/D peptides, RF3 achieves 86% correct chiral center prediction.
  • Interface Accuracy: RF3 shows improved performance on protein-protein and protein-ligand interactions, with scores often falling between AF3 and other open-source models like Boltz. For example, in protein-protein interactions, RF3 (1/24) achieves a median Interface IDDT of 0.61 compared to AF3's 0.66.
  • Templating: Providing ligand conformers or pairwise distances improves protein-ligand interface accuracy from 0.821 to 0.890.
  • AtomWorks Efficiency: The framework allows for rapid parsing of structural files (milliseconds) and efficient data processing pipelines.
  • Code Reuse: AtomWorks enables over 80% code sharing across different models, with LigandMPNN code being reduced from 2,000+ lines to a 100-line declarative pipeline.

Limitations and Future Work

  • ML Force Field Integration: While exploring the integration of ML force fields (like MACE/Egret-1) as atom-level features, the authors found it did not significantly improve accuracy in their preliminary tests. They suggest this area merits further exploration.
  • Chirality Featurization: The authors acknowledge that in practice, they may not always have access to all four substituents of a chiral center (e.g., hydrogen), but they found sufficient information to determine chirality.
  • Future Work: The authors anticipate AtomWorks will accelerate the development of next-generation open-source biomolecular ML models and plan to release the framework, training data, code, and model weights under a permissive BSD license. They also suggest further exploration of ML force field integration.

Summary updated

Abstract

Deep learning methods trained on protein structure databases have revolutionized biomolecular structure prediction, but developing and training new models remains a considerable challenge. To facilitate the development of new models, we present AtomWorks: a broadly applicable data framework for developing state-of-the-art biomolecular foundation models spanning diverse tasks, including structure prediction, generative protein design, and fixed backbone sequence design. We use AtomWorks to train RosettaFold-3 (RF3), a structure prediction network capable of predicting arbitrary biomolecular complexes with an improved treatment of chirality that narrows the performance gap between closed-source AlphaFold3 (AF3) and existing open-source implementations. We expect that AtomWorks will accelerate the next generation of open-source biomolecular machine learning models and that RF3 will be broadly useful as a structure prediction tool. To this end, we release the AtomWorks framework (https://github.com/RosettaCommons/atomworks), together with curated training data, code and model weights for RF3 (https://github.com/RosettaCommons/modelforge) under a permissive BSD license.
BibTeX
@misc{corley2025accelerating,
  author = {Nathaniel Corley and Simon V. Mathis and Rohith Krishna and Magnus S. Bauer and Terry Thompson and Woody Ahern and Maxwell Kazman and Rafael I. Brent and Kieran Didi and Andrew Kubaney and Lilian McHugh and Arnav Nagle and Andrew Favor and Meghana Kshirsagar and Pascal Sturmfels and Yanjing Li and J. Laurence Butcher and Qiang Bo and Lars L. Schaaf and Raktim Mitra and Katelyn V. Campbell and Odin Zhang and Roni Weissman and Ian R. Humphreys and Qian Cong and Jonathan Funk and Shreyash Sonthalia and Píetro Lió and David Baker and Frank DiMaio},
  title = {{Accelerating Biomolecular Modeling with AtomWorks and RF3}},
  year = {2025},
  doi = {10.1101/2025.08.14.670328},
  howpublished = {bioRxiv},
}
Figure