BoltzGen Technical Analysis: Universal Binder Design with All-Atom Generative Models

Original: Stark et al., bioRxiv 2025
DOI: 10.1101/2025.11.20.689494

1. Research Background

1.1 Challenges in De Novo Binder Design

De novo binder design holds tremendous potential for automating drug discovery. This task requires models to understand the three-dimensional structural features of target-binder interactions and generate new molecules with specific binding capabilities. This challenge involves complex structural reasoning and exploration of vast sequence-structure spaces.

1.2 Limitations of Existing Methods

Current technologies face several key limitations: First, modality specificity—many methods are optimized only for specific classes of biomolecules (such as nanobodies or peptides). Second, training data similarity bias—existing methods are typically tested on targets with closely related complexes in the training data, while the true value of de novo design lies in its ability to extrapolate beyond simple targets. Third, lack of flexible control over the design process, making it difficult to meet diverse needs in actual discovery activities.

1.3 Need for Unified Design and Structure Prediction

Models primarily learn to simulate physical laws through provided examples, so extending the generality of methods helps improve their design capabilities for specific categories. Unifying design with structure prediction enables models to simultaneously learn folding physics and binding interactions, thereby enhancing structural reasoning capabilities.

2. Technical Architecture

2.1 All-Atom Generative Model Formalism

BoltzGen employs an all-atom generative model that directly models atomic coordinates rather than simplified residue representations. This representation enables the model to capture fine-grained atomic-level interactions, including hydrogen bonds, hydrophobic interactions, and van der Waals forces.

2.2 Architecture Components

The model architecture includes the following core components:

• Geometric Representation: Pure geometry-based residue type representation, supporting simultaneous training of design and folding tasks
• Trunk: The backbone network processing input sequence and structural information
• Denoising Module: Diffusion model for generating atomic coordinates

Unlike previous design models, BoltzGen matches the performance of state-of-the-art folding models, achieving true unification of design and prediction.

2.3 Unification with Structure Prediction

The key innovation lies in integrating design and folding tasks into a single model for simultaneous training. This unification enables the model to:

• Learn mapping from sequence to structure (folding task)
• Learn mapping from target structure to binder sequence-structure (design task)
• Share structural representations and physical knowledge between both tasks

The result is a model that can both accurately predict structures for given sequences and design new binders for given targets.

2.4 Architectural Relationship with Boltz-1/2

BoltzGen builds upon the technical foundations of Boltz-1 and Boltz-2, inheriting their trunk architecture and diffusion generation framework. The main differences include: introduction of geometric representation for flexible residue type design, extension of generation capabilities from structure prediction to de novo design, and integration of a design specification language to support flexible constraints.

3. Design Specification Language

BoltzGen provides a flexible design specification language that allows users to constrain the design process according to specific application requirements.

3.1 Covalent Bond Constraints

Supports specifying covalent bond connections in designs, such as disulfide bonds (covalent bonds between two cysteines). This enables the design of cyclized peptides and other molecules with enhanced stability.

3.2 Structural Constraints

Includes partial structure constraints (specifying partial three-dimensional structures of binders), binding site constraints (specifying desired binding regions on targets), and "non-binding" constraints (specifying regions to avoid interactions).

3.3 Residue Identity Constraints

Allows fixing residue types or sequence patterns at specific positions, such as fixing nanobody framework regions and only designing CDR loops, or preserving known functionally important residues.

3.4 Application Scenario Examples

• Nanobody Design: Randomly select from 4 default scaffolds, fix framework regions, replace 3 CDR regions with loops of random length
• Disulfide Cyclic Peptide Design: Specify two cysteines forming a covalent bond with design residues in between
• Intrinsically Disordered Protein Binding: Specify binding to disordered regions while avoiding interactions with structured regions

4. Wet-Lab Validation

BoltzGen was tested across 8 independent wet-lab validation projects involving multiple collaborating laboratories, each selecting targets and output modalities relevant to their specific applications.

4.1 Nanobody/Protein Design for 9 Novel Targets

Experiments conducted by Adaptyv Bio. Selected 9 targets ensuring no >30% sequence homologous binding structures in PDB. Generated 60,000 nanobodies and 60,000 proteins (length 80-140) per target without specifying binding sites. Validated 15 designs per target experimentally—both nanobodies and proteins achieved a 66% nM-level binder success rate (6/9 targets). All successful designs passed human serum albumin (HSA) specificity screening with no non-specific binding.

4.2 Bioactive Peptide Binder Design

Experiments conducted by UCSF. Targeted 3 antimicrobial and cytotoxic peptides (protegrin: disulfide-rich β-hairpin; melittin: helical upon membrane binding; indolicidin: polyproline II or amphipathic conformation). Tested 6 designs per target—2 achieved nM affinity, 1 achieved µM affinity, with ability to neutralize antimicrobial and hemolytic activity.

4.3 Disordered Region Binding (NPM1)

Experiments conducted by MPI. NPM1-c mutant is a known driver of acute myeloid leukemia. Generated 20,000 peptide designs (length 40-80), using binding site conditioning to target disordered regions while avoiding structured β-sheet regions. Tested top 5 designs—1 reliably localized to nucleoli in living cells, suggesting successful NPM1 binding. This represents in vivo evidence of de novo designed proteins binding disordered proteins in living cells.

4.4 Site-Specific Peptide Design (RagC)

Experiments conducted by IOCB Boston. RagC GTPase is a core component of cellular nutrient sensing pathways. Used one interaction surface of RagC as binding site input, generated 10,000 designs (length 5-20). Tested 29, discovered 7 binders with highest affinity 3.5 µM, second highest 60 µM.

4.5 Disulfide Cyclic Peptides (RagA:RagC)

Experiments conducted by IOCB Boston. Designed disulfide-cyclized peptides targeting RagA:RagC dimer (length 10-18), specifying interaction surface as binding site, two cysteine covalent bonds, 6 design residues in the middle, and 1-5 design residues on each side. Generated 50,000 designs, tested 24, discovered 14 binders, with 8 resolved affinities—highest 80 µM, second highest 164 µM.

4.6 Viral Protein Nanobodies

Experiments conducted by UC Irvine. Selected two recently deposited PDB monomer targets: Penguinpox cGAMP PDE (degrades cyclic dinucleotides to inhibit host STING signaling) and Bordetella FhaB (adhesin protein). Generated 60,000 nanobodies per target, selected 7 each for yeast surface display. Penguinpox discovered 1 binding signal, Hemagglutinin discovered 7 binding signals (affinity up to 2 µM).

4.7 Small Molecule Binding Proteins

Experiments conducted by UCSF. Designed binding proteins for two small molecules: rucaparib (10,000 designs, length 140-180) and rhodamine derivative (20,000 designs). Rucaparib: tested 6, 5 showed binding with 50-150 µM affinity. Rhodamine derivative: tested 4, all showed weak binding with 30-250 µM affinity. In contrast, previous expert-guided specialized methods designed low nM binders for rucaparib.

4.8 Antimicrobial Peptides (GyrA)

Experiments conducted by MIT. Designed inhibitory peptides targeting bacterial DNA gyrase A subunit (GyrA), specifying GyrA self-interaction surface as binding site, generating peptides of length 10-50. Selected 1,808 designs for growth inhibition assays—352 (19.5%) inhibited E. coli growth >4-fold. Mutated 3 residues closest to target to alanine to validate binding mechanism—54 (3.0%) lost activity.

4.9 5 Benchmark Target Experiments

Experiments conducted by Adaptyv Bio. Designed binders for PD-L1, TNFα, PDGFR, IL-7Rα, and InsulinR—targets with known binders in training data. Generated 30,000-60,000 designs per target, specifying literature binding sites. Both nanobodies and proteins achieved 80% nM-level binder success rate (4/5 targets).

5. Limitations and Discussion

5.1 Affinity Range

BoltzGen-designed binders primarily fall in the µM to nM affinity range, not yet reaching the pM levels common for therapeutic antibodies and nanobodies. For example, in the rucaparib case, expert-guided specialized methods achieved low nM binders while BoltzGen only reached 50-150 µM. This indicates room for improvement in affinity optimization for general-purpose design models.

5.2 Expression Success Rate

The paper does not report protein expression success rate information for designed proteins. Expression failures can result from various causes (such as misfolding, hydrophobic patches causing aggregation), representing a key bottleneck for in vitro validation. More expression data would help assess the practical usability of the model.

5.3 Data Availability

Some experimental data remains confidential at collaborators' request, with the paper stating updates will be provided when further results become available. This data incompleteness limits independent validation and comprehensive evaluation possibilities.

5.4 Comparison with Expert-Guided Methods

The rucaparib case highlights the gap between general-purpose models and expert-guided specialized methods. Specialized methods achieve low nM binding by identifying specific chemical groups on small molecules, while BoltzGen as a general-purpose model only reaches moderate µM affinity. This suggests that integration of domain knowledge remains valuable in specific application scenarios.

5.5 Definition of Novel Targets

The paper defines "novel targets" as those with no >30% sequence homologous binding structures in PDB. However, this does not guarantee that target surfaces lack patches suitable for high-affinity binding. Some targets may fundamentally lack the capacity for high-affinity protein-protein or nanobody-protein binding, and the 66% success rate should be interpreted cautiously in this context.

6. Conclusion

BoltzGen represents significant progress in the field of de novo binder design, achieving high-success-rate design across modalities (nanobodies, proteins, peptides, cyclic peptides) within a unified all-atom generative model framework for the first time. The model's 66% nM-level binder success rate on novel targets demonstrates its ability to extrapolate beyond training data. The flexibility of the design specification language enables the model to adapt to diverse practical application requirements.

However, the model has limitations in affinity range (not reaching pM levels), gaps compared to expert-guided methods, and incomplete experimental data. For drug discovery applications, BoltzGen provides a powerful starting-point design platform, but high-affinity optimization and developability improvements still require subsequent engineering.

References:
[1] Stark H, Faltings F, Choi MG, et al. BoltzGen: Toward Universal Binder Design. bioRxiv 2025. https://doi.org/10.1101/2025.11.20.689494