henhaobio
Sign in

bioRxivBioengineeringPreprint

BoltzGen: Toward Universal Binder Design

Hannes StarkiD, Felix FaltingsiD, MinGyu Choi, Yuxin XieiD, Eunsu Hur, Timothy J. O’DonnelliD, Anton BushuieviD, Talip Uçar, Saro Passaro, Weian Mao, Mateo ReveiziD, Roman BushuieviD, Tally PortnoiiD, Tomáš PluskaliD, Josef ŠiviciD, Karsten KreisiD, Arash VahdatiD, Shamayeeta Ray, Jonathan T. Goldstein, Andrew SavinoviD, Jacob A. Hambalek, Anshika GuptaiD, Diego A. Taquiri-Diaz, Yaotian Zhang, Samuel J. Snyder, , Tommi JaakkolaiD

PDFDOICodeData

In-depth summary

Claude Opus 5.5 · xhigh effort · 2 Nov 2026 · from the full text of version 2 (16 Jun 2026) · CC BY 4.0

An all-atom diffusion model unifying biomolecular structure prediction and sequence-structure generation designs experimentally validated nanobodies, mini-proteins, and linear or cyclic peptides across 26 diverse targets.

Prior de novo binder design methods are often restricted to specific biomolecular classes and benchmarked on targets with closely related complexes already present in the Protein Data Bank. BoltzGen unifies all-atom structure prediction and binder design via a continuous 14-atom geometric residue encoding and a flexible design specification language, generalizing across modalities to unseen targets with <30% sequence identity to any bound PDB structure.

Key results

  1. ShownAgainst ten novel protein targets lacking homologous bound structures in the PDB (<30% sequence identity), BoltzGen generated experimental screening hits for 60% of targets with nanobodies and 50% with mini-proteins while testing at most 15 designs per target.

    From 60,000 generated candidates per target (lengths 80-140), testing the top 15 nanobodies and 15 mini-proteins via surface plasmon resonance and biolayer interferometry yielded hits on 6 of 10 and 5 of 10 novel targets (sequence identities 1% to 27.9%), with 0 detectable binding to human serum albumin and 0 off-target hits for an IDI2 mini-protein across >21,000 human proteins on HuProt v4.0. · Abstract; Wetlab Validation; 2.3 Designing Nanobodies and Proteins against 10…; Human Proteome Microarray Selectivity Profiling

  2. ShownTesting six or fewer designs per target produced nanomolar to single-digit micromolar protein binders that neutralize antimicrobial and hemolytic activities of three structurally diverse bioactive peptides and bind three small molecules.

    Binders neutralized Bacillus subtilis inhibition by protegrin (Kd = 1.2 and 7.2 µM), indolicidin (sub-µM to <5 µM; nanomolar by SPR for indo4), and melittin (Kd = 0.41 and 4.4 µM; 1 molar equivalent reversed erythrocyte hemolysis near HD50 = 1.2 µM); small-molecule campaigns yielded binders for heme (Kd = 697 ± 663 nM, 2.40 µM, 2.84 µM), brilacidin (Kdiss <= 3.9 µM), and rucaparib (Kd = 43.0-151.5 µM). · Wetlab Validation; 2.4 Designing Proteins to Bind Bioactive…; 2.5 Proteins Binding and Inhibiting Small…

  3. ShownDe novo peptide campaigns achieved high hit rates for cyclic and linear binders of the RagA/RagC GTPase heterodimer, live-cell engagement of intrinsically disordered regions in NPM1 and NUP98, and specific in-cell inhibition of Escherichia coli GyrA.

    SPR confirmed RagA/RagC binding for 23/26 (88%) cyclic peptides (Kd = 6-629 µM) and 8/27 (~30%) linear peptides (Kd = 11-1073 µM); 3/10 NPM1 IDR binders enriched in U2OS nucleoli (2 confirmed by BLI) and 1/9 NUP98 binders dissolved NUP98::DDX10 condensates; 352/1,808 (19.5%) GyrA peptides inhibited E. coli growth >=4-fold vs 0.45% of eGFP controls (54 or 3.0% strongly interface-specific). · 2.6 Linear and Disulfide Bonded Peptides against…; NPM1 target; NUP98 target; 2.8 Designing Antimicrobial Peptides that…

  4. ShownBoltzGen matches Boltz-2 structure prediction accuracy while generating higher target-conditioned structural diversity than RFdiffusion and RFdiffusionAA.

    Trained jointly on PDB (through 06/01/2023) and AlphaFold2/Boltz-1 distillation sets using a continuous 14-atom geometric encoding and dilated noise schedule (lambda = 8/3 on [0.6, 0.8], N = 300 steps), BoltzGen matched Boltz-2 lDDT on the post-cutoff test set (<40% sequence similarity) and achieved higher Vendi score diversity across refolded designs (RMSD <= 2.5 Å) for 110 monomer and drug-like ligand targets. · Geometric Encoding of Residue Type; Dilated schedule; 4.1 Structure Prediction; 4.2 Computational Binder Design

Methods, data and reproducibility

BoltzGen extends the Boltz-2 PairFormer trunk and atom/token diffusion module by representing designed residues with 14 atoms whose virtual placements on backbone N, C-alpha, C, and O geometrically encode amino acid identity. Candidates are conditioned via a design language, inverse-folded with BoltzIF, refolded with Boltz-2 in complex and apo states, ranked by worst weighted metric rank, and diversity-filtered.

  • PDB experimental structures (up to 06/01/2023) and AlphaFold2/Boltz-1 distillation datasets (monomers, protein-ligand, RNA, protein-DNA) · multiple species · n = PDB + ~5M AFDB monomers + BindingDB/ChEMBL, Rfam v14.9, and JASPAR 2024 sets
  • Novel monomeric PDB targets (<30% sequence identity to any bound PDB protein) and Chai-2 benchmark targets for nanobody and mini-protein design · multiple species (plus Penguinpox cGAMP PDE and Bordetella FhaB) · n = 10 novel targets, 5 benchmark targets, 2 pathogen targets · PDB: 5M2I, 5JDS, 3MJG, 6P50, 4ZXB, 1JQD, 2A1X
  • Bioactive antimicrobial peptides (protegrin-1, indolicidin, melittin) and small molecules (brilacidin, FeDPP/heme, rucaparib) · Bacillus subtilis ATCC 23857, Escherichia coli, and sheep erythrocytes · n = 6 peptide binders/target; 5 brilacidin, 3 heme, and 6 rucaparib binders tested
  • RagA/RagC GTPase heterodimer, NPM1 and NUP98::DDX10 disordered regions in U2OS cells, and E. coli GyrA C-gate interface · human (U2OS cells, ATCC HTB-96) and Escherichia coli BL21(DE3) · n = 27 linear and 26 cyclic Rag peptides; 10 NPM1 and 9 NUP98; 1,808 GyrA peptides · Addgene: 99664, 237650, 194548; PDB: 6RKS
Methods & tools
BoltzGen, Boltz-2, Boltz-1, BoltzIF, AlphaFold3, Chai-1, RFdiffusion, RFdiffusionAA, ProteinMPNN, SolubleMPNN, LigandMPNN, MMseqs2
Software & environment
Open-source MIT License release including training code, inference code, model weights, and design pipeline.
Assay conditions & compute
Crop sizes up to 768 residues for folding and 512 for design; structure prediction test set evaluated on 40 GB GPUs and pipeline runtime benchmarked on a single NVIDIA A100 GPU (filtering takes ~20 seconds on CPU).
Reproducibility gaps
Some experimental validation datasets remain temporarily confidential at collaborators request and are slated for future updates.
Limitations (3)
  • For designed minibinders of length 73-76 amino acids, generation diversity collapses for certain targets because the model memorizes and frequently samples ubiquitin (which appears in >900 or >1000 PDB entries). · stated by the authors
  • Off-target counter-selection and developability constraints are currently applied only during post-generation filtering rather than integrated directly into the diffusion sampling trajectory. · stated by the authors
  • Yeast surface display testing of top-scoring nanobody designs against two novel pathogen targets (Penguinpox cGAMP PDE and Bordetella FhaB) showed only weak binding signals at 1.6-2 µM antigen, precluding EC50 determination without affinity maturation. · stated by the authors

Read this if you want an open-source, all-atom generative pipeline to design nanobodies, mini-proteins, or linear and cyclic peptides against proteins, disordered regions, or small molecules.

Citations
90

Abstract

We introduce BoltzGen, an all-atom generative model for designing proteins and peptides across all modalities to bind a wide range of biomolecular targets. BoltzGen builds strong structural reasoning capabilities about target-binder interactions into its generative design process. This is achieved by unifying design and structure prediction, resulting in a single model that also reaches state-of-the-art folding performance. BoltzGens generation process can be controlled with a flexible design specification language over covalent bonds, structure constraints, binding sites, and more. We experimentally validate these capabilities in eight diverse design campaigns with functional and affinity readouts across 26 targets. In our experiments, binder modalities span from nanobodies to disulfide-bonded peptides, and targets from disordered proteins to small molecules. In particular, we identify nanobody binders for novel targets with low similarity to proteins with already known bound structures. We release model weights, data, and both inference and training code at: https://github.com/HannesStark/boltzgen.

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
THISCho '25Uçar '26Watson '26Team '25Jin '26Yan '25
76% semantic matchMolecular Biology
Yehlin Cho, Martin Pačesa, Zhidian Zhang+2 more·bioRxiv·Apr 6, 2025· 29 citations