How to Use BoltzGen

BoltzGen is a universal all-atom generative model for biomolecular binder design. Instead of separating inverse design from downstream folding, the method unifies binder generation and structure prediction in one diffusion-based framework so sequence proposals are produced together with structurally grounded target-binder complexes.

The paper expands binder design beyond a single modality: proteins, peptides, nanobodies, and cyclic peptides can be generated against protein, nucleic-acid, and small-molecule targets. On Neurosnap, researchers choose a design mode, provide either a target structure or ligand definition depending on the workflow, and use site constraints to guide the search toward experimentally relevant interfaces.

A key idea in BoltzGen is that structural reasoning is built directly into generation. The model operates with an all-atom geometric representation and a flexible design specification language that can encode binding sites, non-binding regions, covalent relationships, and other architectural constraints. That makes it more suitable for difficult binder-design tasks than pipelines that first generate sequences and only later ask whether they can fold or bind.

On Neurosnap, Mode selects among protein, peptide, nanobody, and protein-small-molecule campaigns. Target Chains, Binding Sites, and Non-Binding Sites let researchers narrow the interface they want to explore, while Minimum Length, Maximum Length, Number Designs, Diffusion Batches, Budget, and the cyclic option govern diversity, binder size, and how aggressively the run searches sequence-structure space.

This workflow is useful when a campaign needs more than a single docked pose. BoltzGen generates a designed binder set that can be compared for structural consistency, interface placement, and downstream developability screening, which is why it fits early-stage therapeutic design and multimodal binder discovery.

• Study-fit modes: BoltzGen supports protein, peptide, nanobody, and protein-small-molecule design workflows, which makes it relevant across several binder modalities rather than a single campaign type.
• Unified design and folding: Sequence generation and structure prediction happen in the same model, so proposed binders are evaluated in structural context from the start.
• Constraint-aware design: Target-chain selection, binding-site and non-binding-site specification, length bounds, diversity controls, and cyclic design options help tailor runs to the biological question.
• Research workflow fit: The resulting binder pool is designed for comparison and downstream filtering, which is useful when moving from target definition to experimental candidate selection.