henhaobio
Sign in

bioRxivMolecular BiologyPreprint

Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction

Saro Passaro, Gabriele CorsoiD, Jeremy WohlwendiD, Mateo ReveiziD, Stephan ThaleriD, Vignesh Ram SomnathiD, Noah GetziD, Tally PortnoiiD, Julien Roy, Hannes StarkiD, David Kwabi-Addo, Dominique BeainiiD, Tommi JaakkolaiD, Regina BarzilayiD

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

Abstract

Accurately modeling biomolecular interactions is a central challenge in modern biology. While recent advances, such as AlphaFold3 and Boltz-1, have substantially improved our ability to predict biomolecular complex structures, these models still fall short in predicting binding affinity, a critical property underlying molecular function and therapeutic efficacy. Here, we present Boltz-2, a new structural biology foundation model that exhibits strong performance for both structure and affinity prediction. Boltz-2 introduces controllability features including experimental method conditioning, distance constraints, and multi-chain template integration for structure prediction, and is, to our knowledge, the first AI model to approach the performance of free-energy perturbation (FEP) methods in estimating small molecule-protein binding affinity. Crucially, it achieves strong correlation with experimental readouts on many benchmarks, while being at least 1000x more computationally efficient than FEP. By coupling Boltz-2 with a generative model for small molecules, we demonstrate an effective workflow to find diverse, synthesizable, high-affinity binders, as estimated by absolute FEP simulations on the TYK2 target. To foster broad adoption and further innovation at the intersection of machine learning and biology, we are releasing Boltz-2 weights, inference, and training code 1 under a permissive open license, providing a robust and extensible foundation for both academic and industrial research.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (18 Jun 2025) · CC BY 4.0

Boltz-2 predicts protein–ligand affinity from its co-folded structures, nearing free-energy perturbation accuracy (average Pearson 0.66 on a 4-target FEP+ subset) at over 1000× lower compute.

Free-energy perturbation (FEP) is the most accurate in-silico way to rank binding affinity but is too slow for large-scale use, docking is too imprecise, and the authors note that no AI model had matched FEP. Boltz-2 puts an affinity module on top of co-folding representations, trains it on millions of standardized assay measurements, and extends Boltz-1 with ensemble training and user controls such as method, template and contact conditioning.

Key results

  1. ShownOn the two FEP+ benchmark subsets (OpenFE and a 4-target set), Boltz-2 outperformed inexpensive physics-based methods and ML baselines and came close to free-energy simulation accuracy at a fraction of the compute.

    Average Pearson 0.66 on a 4-target FEP+ subset that also has absolute FEP (ABFE) and Fragment Molecular Orbital baselines; close to OpenFE relative FEP on the 876-measurement OpenFE set; more than 1,000× faster than FEP. Training proteins with ≥ 90% sequence identity to FEP+ proteins were removed. FEP correlations themselves are not given in this text. · Boltz-2 approaches FEP accuracy on public…; Abstract

  2. ShownWithout fine-tuning, Boltz-2 beat every top-ranked entry in the CASP16 blind affinity challenge, but on blinded industrial assays its accuracy was uneven.

    CASP16: 140 protein–ligand pairs across two targets, evaluated retrospectively; the margin is called clear but is not quantified in this text. Eight internal Recursion assays: well ahead of other ML baselines, yet Pearson > 0.55 on only 3 of 8 and limited on the other 5. · Boltz-2 approaches FEP accuracy on public…; Introduction; Understanding the range of applicability of the…

  3. ShownIn retrospective virtual screening on MF-PCBA, Boltz-2 retrieved actives from large, imbalanced libraries much better than prior ML models, ipTM-based ranking and docking.

    Average precision nearly doubled relative to these baselines and the enrichment factor reached 18.4 at a 0.5% threshold; AUROC was also assessed, but its values are not given in this text. · Boltz-2 enables accurate large-scale virtual…

  4. SuggestedIn a prospective TYK2 screen judged by absolute FEP simulations rather than experiments, Boltz-2's picks from Enamine libraries and its SynFlowNet-generated molecules were mostly predicted binders.

    Boltz-ABFE predicted binding for 8 of the top 10 Hit Locator Library picks (460,160 compounds screened), all 10 Kinase Library picks (64,960 screened) and none of 10 random compounds; all 10 de novo SynFlowNet candidates were predicted binders with higher mean affinity, using 117k vs 460k Boltz-2 evaluations. · Boltz-2 enables accurate large-scale virtual…

  5. ShownStructure prediction matched or modestly beat Boltz-1 (largest gains for RNA and DNA–protein complexes, a moderate gain for antibody–antigen), edged Chai-1 and ProteinX and trailed AlphaFold3 slightly; MD-conditioned ensembles were competitive with BioEmu and AlphaFlow, generally tracking simulated RMSF more closely.

    PDB complexes from 2024 and 2025 dissimilar to all models' training data; Boltz-steering (Boltz-2x) improved physicality metrics. On Polaris-ASAP protease ligand poses, Boltz-2 beat Boltz-1 and the top challenge entries without fine-tuning. On held-out mdCATH and ATLAS clusters, its MD ensembles generally correlated better with simulated RMSF than Boltz-1, BioEmu and AlphaFlow. · PDB evaluation set; Antibody benchmark; Polaris-ASAP challenge (SARS-CoV-2 and MERS-CoV); Boltz-2 can better capture local protein…

Methods, data and reproducibility

Boltz-2 co-folds proteins, nucleic acids and small molecules (trunk, denoising module with optional physics steering, confidence module), trained on PDB structures, NMR and molecular dynamics ensembles and distillation sets; a separate affinity module, a PairFormer over the predicted pocket with binding-likelihood and log-affinity heads, is trained afterwards on curated binary and continuous assay data.

  • Training structures: PDB entries released before 2023-06-01, plus distillation from AlphaFold2 monomers and high-confidence Boltz-1 complexes (RNA, protein–DNA, ligand–protein, MHC)
  • Molecular dynamics ensembles from MISATO, ATLAS and mdCATH (held-out mdCATH and ATLAS clusters used for evaluation)
  • Affinity training data from ChEMBL, BindingDB and PubChem, PubChem HTS assays, a CeMM fragment screen and MIDAS, plus synthetic decoys · n = millions of standardized assay measurements
  • Structure evaluation: PDB complexes from 2024 and 2025, an antibody–antigen benchmark, and Polaris-ASAP ligands bound to SARS-CoV-2 and MERS-CoV main proteases
  • FEP+ benchmark: OpenFE subset and a 4-target subset with ABFE and FMO baselines · n = 876 measurements (OpenFE subset); 4 targets
  • Blind affinity sets: CASP16 affinity challenge and blinded internal Recursion assays · n = 140 protein–ligand pairs across two targets (CASP16); eight assays (Recursion)
  • MF-PCBA retrospective virtual-screening benchmark
  • TYK2 prospective screen: Enamine Hit Locator Library, Kinase Library and SynFlowNet sampling of Enamine REAL space · n = 460,160 and 64,960 compounds; 76B REAL space
Methods & tools
Boltz-2, Boltz-2x (Boltz-steering), Boltz-1, AlphaFold3, Chai-1, ProteinX, AlphaFold2, AlphaFlow, BioEmu, SynFlowNet (GFlowNet), Boltz-ABFE, OpenFE
Assay conditions & compute
Reported as at least 1000× more efficient than FEP; bfloat16 mixed precision and the trifast triangle-attention kernel reduce runtime and memory, allowing 768-token training crops. No hardware, run times or memory figures are given in this text.
Reproducibility gaps
The paper says Boltz-2 weights, inference and training code are released under a permissive open licence, but no link is in this text; training data come from public databases named without URLs, and the Recursion assays are internal.
Limitations (6)
  • Affinity predictions depend on the predicted pocket and interface being right; cofactors such as ions, water and multimeric partners are not handled explicitly, and the affinity crop can truncate long-range or allosteric interactions. · stated by the authors
  • Affinity accuracy varies strongly between assays (Pearson > 0.55 on only 3 of 8 Recursion assays), and whether this stems from structure errors or poor generalization to new protein families or ligands is unresolved. · stated by the authors
  • TYK2 hits were checked only by a single repeat of the authors' Boltz-ABFE pipeline, not by experiments, and may look optimistic because Boltz-2 already does well on this target (Pearson R = 0.83). · stated by the authors
  • Structure accuracy stays close to its predecessors, with failures on large binding-induced conformational changes and large complexes; MD ensembles are less diverse than BioEmu and AlphaFlow and came from a relatively small MD training set. · stated by the authors
  • The affinity head is trained on mixed Ki, Kd and IC50-type measurements, so its output is a ranking-oriented, roughly IC50-like score rather than a calibrated binding free energy. · stated by the authors
  • The CASP16 comparison is retrospective, and the text documents leakage filtering (≥ 90% sequence identity) only for the FEP+ benchmark, not for CASP16 or MF-PCBA targets. · noted in this summary

Read this if you rank protein–ligand affinities or run structure-based virtual screens and want an open co-folding model benchmarked against FEP, docking and ML baselines.

Citations
710

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
THISTilewale '26Amini '26Chu '26Piao '26Gazizov '23Kaminow '26
80% semantic matchBioinformatics
Atharva Tilewale, Dhaval Patel·bioRxiv·Sep 24, 2026