henhaobio
Sign in

bioRxivBiochemistryPreprint

Accelerating Biomolecular Modeling with AtomWorks andRF3

Nathaniel CorleyiD, Simon V. MathisiD, Rohith KrishnaiD, Magnus S. BaueriD, Tuscan R. Thompson, Woody AherniD, Maxwell W. KazmaniD, Rafael I. BrentiD, Kieran DidiiD, Andrew KubaneyiD, Ryan McHughiD, Arnav NagleiD, Andrew FavoriD, Meghana KshirsagariD, Pascal Sturmfels, Yanjing LiiD, Jasper ButcheriD, Bo QiangiD, Lars L. SchaafiD, Raktim MitraiD, Katelyn V. CampbelliD, Odin ZhangiD, Roni Weissman, Ian R. HumphreysiD, Qian CongiD, , Frank DiMaioiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

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.

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (15 Aug 2025) · CC BY 4.0

AtomWorks, a modular atom-level data framework, trains structure prediction and design models; the resulting RF3 improves chirality handling and performs between AF3 and Boltz.

Open-source biomolecular modeling is slowed by heterogeneous PDB data and bespoke per-model pipelines that duplicate effort. AtomWorks offers a shared atom-level framework, and RF3 narrows the gap between closed-source AF3 and open-source models.

Key results

  1. ShownRF3 predicts ligand chirality well without inference-time guidance, significantly outperforming Boltz-2 and scoring numerically above AF3, though that gap is not significant. It uses a learned geometric chirality feature plus a chirality-inversion augmentation.

    Correct chirality at 88% of ligand chiral centers for RF3 vs 84% AF3 and 76% Boltz-2; p < 0.001 (RF3 vs Boltz-2), p = 0.002 (AF3 vs Boltz-2), p = 0.064 (RF3 vs AF3). · RF3 accurately adheres to specified…; Figure legends

  2. SuggestedOn held-out cyclic mixed L/D peptides, RF3 gets most chiral centers right and has lower backbone RMSD than AF3 and Boltz-1x. Boltz-1x with guidance satisfies all centers by construction.

    86% of chiral centers correct (figure legend says 85%) vs 70% AF3 and 100% Boltz-1x with guidance; mean backbone RMSD 1.74 Å vs 1.88 AF3 and 2.24 Boltz-1x; 11 PDB structures. · RF3 accurately adheres to specified…; A.5.3 Mixed Chirality L/D Macrocyclic Peptide…; Figure legends

  3. ShownAtom-level conditioning lets users supply ligand conformers or protein-ligand distances. This raises protein-ligand interface accuracy, and the model keeps to the supplied ligand conformation.

    Median protein-ligand interface accuracy rises from 0.821 to 0.882 (rigid ligand folding) and from 0.821 to 0.890 (holo distance constraints); templated ligand-only lDDT 0.991 (129 complexes). · RF3 enables flexible user control through…; Figure legends

  4. ShownRF3 generally lands between AF3 and the open-source Boltz models. Extending the training cutoff to 1/2024 improves it, and on antibody-antigen complex prediction it again falls between AF3 and the other open models.

    Median interface lDDT, earlier cutoff vs 1/2024 cutoff: protein-protein 0.571 vs 0.607; protein-ligand 0.766 vs 0.798; protein-DNA 0.415 vs 0.523. Antibody DockQ > 0.23: RF3 33%, AF3 44%, Boltz-2 22%, Chai-1 28%. · RF3 narrows the performance gap between existing…; A.5.2 Antibody Evaluation Set; Figure legends

  5. ClaimedThe framework lets most code be shared across models and gives lower-energy reference conformers than another open-source model. These are the authors' own claims on their own pipelines, with no external comparison of development effort.

    Over 80% of code shared across networks; 2,000+ lines of LigandMPNN code replaced by a 100-line pipeline; median reference conformer energy 67 vs 109 kJ/mol for another open model; test coverage over 85%. · AtomWorks prioritizes high-quality data; AtomWorks enables the scalable training of…; AtomWorks follows industry-grade testing and…

Methods, data and reproducibility

AtomWorks represents structures as Biotite AtomArrays processed by chained Transforms. RF3 is an AF3-style diffusion model trained in three stages on PDB plus distillation sets, with geometric chirality features. It is evaluated on a recent low-homology PDB set, an antibody-antigen set, and a 2022 macrocycle set.

  • PDB training data with date cutoffs of 9/2021 and 1/2024, plus AFDB domain-pair and monomer distillation sets and a disordered-region distillation set
  • New RNA distillation set built from RNAcentral (January 2025) sequences, with structures predicted by an earlier RF3 version · n = 10K examples
  • New protein-nucleic acid complex distillation set (cis-BP, UniProbe, TRANSFAC, HT-SELEX, ChIP-Seq sources) · n = 27K examples
  • Recent PDB evaluation set, released 2024-01-01 to 2025-07-13, low-homology chains and interfaces, clustered
  • Antibody-antigen test set from SAbDab, de-leaked against training at 40% antigen sequence identity · n = 75 structures (58 scored for all methods; 36 antigen clusters)
  • Mixed-chirality cyclic L/D peptide set from a Cell 2022 macrocycle design paper · n = 11 structures · PDB: 7ubc, 7ubd, 7ube, 7ubf, 7ubg, 7ubh, 7ubi, 7uzl, 8cto, 8cun, 8cwa
  • Protein-ligand complexes used for ligand-conformer templating tests · n = 129 complexes
Methods & tools
AtomWorks, RosettaFold-3 (RF3), Biotite AtomArray, RDKit, AlphaFold3 (AF3), Boltz-1x / Boltz-2, Chai-1, PoseBusters, MMseqs2 / MMseqs2-GPU, HHBlits, EternaFold, ProteinMPNN / LigandMPNN
Assay conditions & compute
Evaluations ran on NVIDIA A100 GPUs; ProteinMPNN and LigandMPNN were each trained on one A100 GPU for three days.
Reproducibility gaps
Benchmark scripts are promised with the public code release. Curated training data and weights are said to be released, but no data URL or accession is printed. The authors expect some discrepancies from AF3 training because not all AF3 details were published.
Limitations (6)
  • The RF3 training procedure may differ from the one used for the released AF3 weights, because not all AF3 training details were published. · stated by the authors
  • The authors did not ablate HHBlits (pre-2021) versus MMSeqs2-GPU (post-2021) MSAs for training, so any effect of MSA source is unknown. · stated by the authors
  • RF3 performs below AF3 in almost all categories the authors compare, for example antibody-antigen DockQ > 0.23 (33% vs 44%), so a gap to AF3 remains. · stated by the authors
  • Egret-1 force-field embeddings did not significantly improve accuracy over standard conformer embedding in a preliminary test. · stated by the authors
  • The main benchmark takes the best of 5 samples per metric, which favors methods with diverse samples. The antibody (58 scored structures) and macrocycle (11 structures) sets are small and reported without confidence intervals. · noted in this summary
  • The mixed-peptide chirality accuracy is given as 86% in the text and 85% in the figure legend. The text also gives no ablation separating the learned chirality feature from the 2% inversion augmentation. · noted in this summary

Read this if you want a reusable atom-level data pipeline for training structure prediction or design models, or an open AF3-style model with learned chirality and conformer conditioning.

Citations
44

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISYu '25Team '25Xu '23Krishna '23Elliott '24Deng '26
68% semantic matchMolecular Biology
Jingcheng Yu, Zhaoming Chen, Zhaoqun Li+4 more·bioRxiv·Mar 25, 2025