How to Use Boltz-2 (AlphaFold3)

Boltz-2 is an open AlphaFold3-class biomolecular modeling system that predicts all-atom complex structures and estimates binding affinity in the same framework. The paper positions it as a step beyond structure-only co-folding by coupling high-quality multimolecule geometry with quantitative affinity heads, especially for protein-ligand systems where both pose credibility and interaction strength matter.

On Neurosnap, researchers can combine Input Sequences, Input Molecules, cyclic or modified biopolymers, and optional pocket or covalent restraints in a single job. The service is suited to protein-protein, protein-nucleic-acid, and protein-ligand questions where structural plausibility and interaction potency both influence which hypotheses move forward. For complexes, Neurosnap also calculates additional interface-evaluation metrics such as ipSAE, LIS, pDockQ, and pDockQ2 to strengthen intermolecular interaction review beyond the core model confidences.

Methodologically, Boltz-2 extends the Boltz family with a dedicated affinity-prediction head trained on curated biochemical assay data, a deeper PairFormer stack, and ensemble-style supervision from NMR and molecular-dynamics-derived structures to better capture conformational variation. The model also broadens the input representation with cyclic and modification flags, bond-order information, and conditioning mechanisms for pockets, templates, and distance restraints.

The paper and service both emphasize physics-aware steering potentials and flexible sampling. On Neurosnap, options such as Use Inference Time Potentials, Molecular Weight Correction, MSA Mode, Custom MSA, recycling, structure-sampling settings, and separate affinity-sampling settings let users decide whether a run should favor speed, structural fidelity, or more stable affinity estimates. These are platform workflow controls, not separate scientific models.

Researchers typically interpret Boltz-2 through two lenses: geometry and potency. Confidence metrics such as pLDDT, PAE, PDE, ipTM, ipSAE, LIS, pDockQ, and pDockQ2 indicate whether a predicted complex is structurally credible and whether the intermolecular interface looks coherent, while the affinity outputs help rank ligand-containing hypotheses when several poses or chemotypes compete.

• Study-fit inputs: Boltz-2 accepts proteins, nucleic acids, ligands, cyclic polymers, residue modifications, and optional restraints, which keeps multimolecule prediction close to the actual experimental system.
• Joint structure and affinity modeling: The method does not stop at a pose; it combines AlphaFold3-class complex prediction with explicit affinity estimation for more decision-relevant ranking.
• Constraint-aware workflow: Custom MSAs, pocket restraints, covalent restraints, steering potentials, and separate sampling controls give researchers practical ways to tune the run for difficult systems.
• Decision-ready interpretation: Structural confidence summaries, including ipSAE, LIS, pDockQ, and pDockQ2 for complexes, can be reviewed together with affinity estimates when choosing among competing poses, ligands, or assembly hypotheses.