How to Use Boltz-2 (AlphaFold3)

Boltz-2 is an open AlphaFold3-class biomolecular structure-prediction model for all-atom complexes that also estimates small-molecule binding affinity. That combination makes it useful when a project needs more than a docked pose: the same workflow can evaluate whether a protein-protein, protein-nucleic-acid, or protein-ligand assembly is geometrically plausible and, for ligand-containing systems, whether the interaction is ranked as stronger or weaker across candidate chemotypes.

On Neurosnap, researchers can combine Input Sequences, Input Molecules, cyclic or modified biopolymers, optional custom MSAs, and optional Pocket Restraints, Covalent Restraints, or Contact Restraints in one no-code job. The platform returns the model confidences together with additional interface-focused metrics such as ipSAE, LIS, pDockQ, and pDockQ2, which helps determine whether a predicted assembly is globally credible, locally informative, or still too uncertain for downstream use.

Methodologically, Boltz-2 extends the Boltz family with a stronger multimolecular structure model and a dedicated affinity head trained for protein-ligand ranking tasks. The model operates in an all-atom setting across proteins, nucleic acids, and ligands, while also supporting conditioning inputs such as cyclic polymers, residue modifications, templates, pockets, covalent bonds, and explicit contact or distance-style restraints. This matters in practice because many experimental systems are not unconstrained folding problems; they already come with partial mechanistic knowledge from mutagenesis, structural biology, medicinal chemistry, or prior docking.

On Neurosnap, these capabilities map cleanly to researcher-facing controls. MSA Mode and Custom MSA control how evolutionary context is supplied. Recycling and sampling parameters govern the tradeoff between runtime, refinement depth, and structural diversity. Use Inference Time Potentials enables the Boltz steering potentials that are designed to improve physical plausibility, while Pocket Restraints, Contact Restraints, and Covalent Restraints let users encode experimentally motivated interface hypotheses directly into the prediction setup instead of handling them in a separate pipeline.

Researchers should interpret Boltz-2 through two related lenses: structure quality and interaction prioritization. pLDDT, PAE, PDE, pTM, ipTM, ipSAE, LIS, pDockQ, and pDockQ2 are useful for deciding whether chain placement, interface geometry, and local regions are trustworthy enough for mechanistic analysis or follow-up design. For ligand-containing systems, the affinity outputs add a second ranking signal that is better used comparatively across active candidates, analog series, or competing poses than as a standalone experimental surrogate.

• Multimodal biomolecular input: Boltz-2 accepts proteins, nucleic acids, ligands, cyclic polymers, and inline CCD-style modifications in one workflow, which better matches real complex-modeling studies than protein-only predictors.
• Joint pose and potency reasoning: The workflow combines AlphaFold3-class all-atom complex prediction with explicit affinity estimation, which is useful when pose plausibility and ligand ranking both matter.
• Constraint-aware setup: Pocket Restraints, Contact Restraints, Covalent Restraints, templates, steering potentials, and custom MSAs give researchers practical ways to inject prior knowledge into difficult prediction problems.
• Experiment-facing interpretation: Confidence summaries and interface metrics can be reviewed alongside affinity outputs when prioritizing complexes for docking follow-up, mutagenesis, medicinal chemistry, or wet-lab validation.