How to Use Boltz-1 (AlphaFold3)

Boltz-1 is a fully open AlphaFold3-class model for predicting all-atom structures of proteins, nucleic acids, ligands, and mixed biomolecular complexes. The paper emphasizes open access together with strong benchmark performance, making the method relevant for researchers who need state-of-the-art multimolecule structure prediction without depending on closed platforms.

On Neurosnap, a single run can combine Input Sequences, Input Molecules, cyclic biopolymers, residue modifications, and optional pocket or covalent restraints. That makes the service suitable for protein-protein, protein-ligand, nucleic-acid, and mixed-complex questions where prior interaction knowledge may or may not be available.

Boltz-1 follows the all-atom diffusion paradigm introduced for AlphaFold3-class modeling, but the paper highlights several additions that materially affect real-world use. These include dense taxonomy-based MSA pairing for complexes, unified cropping that bridges spatial and contiguous training views, and pocket conditioning designed to remain useful even when only part of an interaction site is specified.

The confidence model is also strengthened relative to simpler heads. Boltz-1 uses a trunk-like confidence network initialized from the main trunk and augmented with features from the reverse diffusion trajectory, which helps explain why confidence summaries such as pLDDT, PAE, PDE, and interchain interface metrics are central to model review on Neurosnap.

On Neurosnap, Model Version, MSA Mode, recycling, sampling controls, and optional pocket or covalent restraints are the main levers. The original Boltz-1 and the newer Boltz-1x option should be understood as platform workflow choices that let users trade off speed, diversity, and physical plausibility depending on whether the goal is exploratory co-folding or higher-quality complex poses.

• Study-fit inputs: Boltz-1 accepts proteins, nucleic acids, ligands, cyclic polymers, residue modifications, and optional restraints, which keeps multimolecule prediction close to the experimental system.
• Multimodal modeling: The method brings AlphaFold3-class all-atom prediction to mixed biomolecular assemblies rather than only single proteins or simple docking setups.
• Constraint-aware runs: Pocket and covalent restraints make it practical to inject prior knowledge when unconstrained co-folding is too ambiguous for the biological question.
• Confidence-guided interpretation: pLDDT, alignment-error, distance-error, and interface-confidence summaries help researchers decide whether a predicted complex is globally reliable or only locally informative.