Boltz-1: Democratizing Biomolecular Interaction Modeling
Jeremy WohlwendiD, Gabriele CorsoiD, Saro Passaro, Noah GetziD, Mateo ReveiziD, Ken Leidal, Wojtek Swiderski, Liam Atkinson, Tally PortnoiiD, Itamar ChinniD, Jacob SilterraiD, Tommi JaakkolaiD, Regina BarzilayiD
Abstract
Understanding biomolecular interactions is fundamental to advancing fields like drug discovery and protein design. In this paper, we introduce BO_SCPLOWOLTZC_SCPLOW-1, an open-source deep learning model incorporating innovations in model architecture, speed optimization, and data processing achieving AO_SCPLOWLPHAC_SCPLOWFO_SCPLOWOLDC_SCPLOW3-level accuracy in predicting the 3D structures of biomolecular complexes. BO_SCPLOWOLTZC_SCPLOW-1 demonstrates a performance on-par with state-of-the-art commercial models on a range of diverse benchmarks, setting a new benchmark for commercially accessible tools in structural biology. Further, we push the boundary of capabilities of these models with BO_SCPLOWOLTZC_SCPLOWO_SCPCAP-C_SCPCAPO_SCPLOWSTEERINGC_SCPLOW, a new inference time steering technique that is able to fix hallucinations and non-physical predictions from the models. By releasing the training and inference code, model weights, datasets, and benchmarks under the MIT open license, we aim to foster global collaboration, accelerate discoveries, and provide a robust platform for advancing biomolecular modeling.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 529
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 4 (6 May 2025) · CC BY 4.0
Open-source Boltz-1 predicts biomolecular complex structures comparably to AlphaFold3 and Chai-1 on CASP15 and recent PDB targets; Boltz-steering lets 97% of test-set poses pass physical checks.
Leading all-atom complex predictors such as AlphaFold3 and Chai-1 ship under exclusive commercial licenses without training code, and machine-learned structures often violate physics through clashes, wrong chirality or overlapping chains. Boltz-1 reaches comparable accuracy while releasing training and inference code, weights, data splits and benchmarks under the MIT license, and adds inference-time steering that repairs most such defects.
Key results
ShownBoltz-1 performs comparably to AlphaFold3 and Chai-1 on all-atom LDDT, DockQ success, protein-ligand LDDT-PLI and ligand RMSD; interface differences sit within confidence intervals, and AlphaFold3's slight mean-LDDT edge is ascribed to extra distillation data aiding RNA/DNA complexes.
Evaluated on 66 CASP15 and 541 test-set structures (5 samples each; 200 sampling steps, 10 recycling rounds, shared MSAs up to 16384 sequences; Chai-1 via chai_lab 0.2.1). AlphaFold3 slightly higher on test-set DockQ > 0.23 success; AlphaFold3 and Boltz-1 slightly above Chai-1 on LDDT-PLI and ligand RMSD < 2 Å; all within confidence intervals. · Benchmarks; Baselines; Evaluation criteria; Results
ShownBoltz-steering (Boltz-1x), which tilts reverse diffusion with flat-bottom physics potentials plus gradient guidance and periodic resampling, removes most hallucinated chain overlaps and stereochemical or steric errors without degrading accuracy.
On the test set, 57% of Boltz-1 top-1 poses fail at least one PoseBusters-derived check (ligand bonds, angles, internal clashes, chirality, stereochemistry, inter-chain clashes); around 27% (Chai-1) and 58% (AlphaFold3) of complexes pass every check, versus 97% for Boltz-1x at similar accuracy. · 4.1 Introduction; 4.2 Method; Evaluation criteria; Results
ShownA new homology- and time-split PDB benchmark is released: validation and test complexes exclude training protein clusters at 0.4 sequence identity and, when ligands are present, require one with Tanimoto similarity of 0.8 or less to training ligands.
From 1728 filtered candidates, stratified selection kept 126 RNA/DNA, 330 ligand or ion, 231 multimer (50% subsampled) and 57 monomer (30% subsampled) structures, 744 in total, then 553 validation structures after a 1024-residue cap; the post-2023-01-13 test set has 593 structures of 100 to 2000 residues. · 2.2 Validation and test sets curation; Evaluation criteria
ClaimedBoltz-1 reached AlphaFold3-level accuracy with a far shorter training schedule than AlphaFold3, a saving the authors partly credit to their architectural and data innovations without isolating each contribution.
68k steps at batch size 128 (53k steps at 384 tokens/3456 atoms sampling PDB and ~270K OpenFold distillation structures equally, then 15k PDB-only steps at 512 tokens/4608 atoms) versus nearly 150k steps at batch size 256 for AlphaFold3, estimated at approximately four times the computing time. · Structure prediction training pipeline; 3 Modeling
SuggestedAblations show accuracy rising roughly monotonically with recycling rounds and diffusion sampling steps, with gains largely plateauing beyond 3 recycling rounds and 50 diffusion steps.
Table 1 ablation on the test set with 5 samples per target, scored as oracle and top-1, using pre-computed MSAs capped at 4,096 sequences; authors warn the metrics are noisy and differ slightly from the main benchmark owing to MSA settings and the subset of structures completing all ablations. · Results; Tables
Methods, data and reproducibility
AlphaFold3-style all-atom diffusion model (MSA/PairFormer trunk plus atom- and token-level denoising transformer) with taxonomy-based dense MSA pairing, unified spatial/contiguous cropping, partial pocket conditioning, Kabsch-aligned reverse diffusion and a trunk-initialized confidence model; Boltz-1x adds Feynman-Kac steering with physics potentials.
- PDB biological assemblies (Assembly 1 from mmCIF) released before 2021-09-30 at resolution of at least 9 Å, cleaned following AlphaFold3, used for training
- OpenFold distillation structures with their provided MSAs, sampled equally with PDB for the first 53k training steps · n = approximately 270K structures
- Validation split: PDB entries released 2021-09-30 to 2023-01-13, resolution below 4.5 Å, unseen protein clusters, dissimilar ligands, at most 1024 residues · n = 553 structures
- Test split: PDB entries released after 2023-01-13 with 100 to 2000 residues, filtered for homology and ligand similarity against all earlier PDB entries · n = 593 structures (541 benchmarked)
- CASP15 targets (not cancelled, with PDB ground truth, matching stoichiometry, below 2000 residues) · n = 76 structures (66 benchmarked)
- Methods & tools
- AlphaFold3 (baseline and architectural starting point), Chai-1 via chai_lab 0.2.1 (baseline), colabfold_search / MMseqs2 (uniref30_2302, colabfold_envdb_202108), MMseqs2 easy-cluster (min-seq-id 0.4) for splits, RDKit (ETKDGv3 conformers, distance-bounds matrices), Chemical Component Dictionary (CCD), UniProt taxonomy annotations, OpenStructure 2.8.0 (LDDT, LDDT-PLI, DockQ, ligand RMSD), PoseBusters physical-validity checks, Trifast Triton kernel for triangle self-attention, OpenFold (chunking scheme, distillation set), Feynman-Kac steering with Sequential Monte Carlo and universal guidance
- Data & accessions
- https://github.com/jwohlwend/boltz
- Software & environment
- Chai-1 run via chai_lab 0.2.1; MSAs from colabfold_search (uniref30_2302, colabfold_envdb_202108); RDKit ETKDGv3 conformers; metrics in OpenStructure 2.8.0; Trifast written in Triton; code, weights and data under the MIT license.
- Assay conditions & compute
- Benchmarks ran on A100 80GB GPUs, excluding structures that ran out of memory; training used 68k steps at batch size 128, and project GPUs came largely from Genesis Therapeutics and the US Department of Energy via NERSC (GenAI@NERSC).
- Reproducibility gaps
- Training GPU count and wall-clock time are not given; benchmark inputs, outputs and scripts are obtained via download instructions in a GitHub repository the authors plan to update regularly, with no archived accession or pinned version in the text.
Limitations (6)
- Architectural changes were screened on a smaller model at various development stages; the authors expect the gains to carry to full-size Boltz-1 but provide no direct ablations at that scale. · stated by the authors
- Unsteered Boltz-1 hallucinates chains stacked on one another and yields steric clashes, strained bonds and angles, wrong chirality or stereobonds and non-planar aromatic rings; 57% of its top-1 test-set poses fail the physical checks. · stated by the authors
- AlphaFold3 keeps a slight edge in mean all-atom LDDT, which the authors link to its extra distillation datasets improving predictions for RNA- and DNA-containing complexes. · stated by the authors
- Recycling and diffusion-step ablations used MSAs capped at 4,096 sequences, are acknowledged as noisy, and differ slightly from the main benchmark, so small non-monotonic effects should not be overinterpreted. · stated by the authors
- Steering potentials directly target chirality, stereochemistry, ligand distance bounds and inter-chain clashes, largely the properties the PoseBusters checks score, so the 97% pass rate is partly by construction; the added inference cost of steering is not quantified. · noted in this summary
- Test-set comparisons exclude complexes with covalently bound ligands (Chai-1 cannot specify them), and both benchmarks drop any structure that failed or ran out of memory for any method on A100 80GB GPUs, so parity on covalent test-set ligands and the largest complexes is unassessed. · noted in this summary
Who should read:Read this if you need an open, retrainable AlphaFold3-class predictor for protein, nucleic-acid and ligand complexes, or want physics-based inference-time steering and a homology-controlled PDB benchmark.
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
- Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction
Saro Passaro, Gabriele Corso, … Regina Barzilay·bioRxiv·Jun 14, 2025
- BoltzDesign1: Inverting All-Atom Structure Prediction Model for Generalized Biomolecular Binder Design
Yehlin Cho, Martin Pačesa, … Sergey Ovchinnikov·bioRxiv·Apr 6, 2025
- FoldBench: An All-atom Benchmark for Biomolecular Structure Prediction
Sheng Yong Xu, Qiantai Feng, … Siqi Sun·bioRxiv·May 22, 2025
- Accelerating Biomolecular Modeling with AtomWorks andRF3
Nathaniel Corley, Simon V. Mathis, … Frank DiMaio·bioRxiv·Aug 14, 2025
- IntelliFold-2: Surpassing AlphaFold 3 via Architectural Refinement and Structural Consistency
Lifeng Qiao, Yan He, … Siqi Sun·bioRxiv·Feb 14, 2026
- Boltz2-Notebook: An Interactive Google Colab Platform for Diffusion-Based Biomolecular Structure and Binding Affinity Prediction using the Boltz2 model.
Atharva Tilewale, Dhaval Patel·bioRxiv·Sep 24, 2026