Boltz-2: Structure & Affinity in One Run

A practical guide to running Boltz-2 and interpreting structure and affinity scores.

Summary

Boltz-2 represents a shift in computational biology, delivering AlphaFold-3 level accuracy in an open, physically aware framework that supports a full spectrum of biomolecular interactions. While many modern models excel at static structural inference, Boltz-2 is engineered to bridge the gap between structure prediction and thermodynamic characterization. It can solve complex heterogeneous systems such as co-folded proteins, DNA, RNA, non-canonical residues, and small molecule ligands in a single inference run.

Beyond structural prediction, Boltz-2 differentiates itself through affinity-oriented outputs for ligand–protein interactions, helping bridge docking and lead optimization. Unlike standard folding tools that report only geometric confidence, Boltz-2 can provide predicted affinity signals such as IC50 and ΔG (kcal/mol), alongside a binder probability score that helps separate likely binders from decoys. These outputs are designed for rapid, high-throughput prioritization of compounds with thermodynamic context, enabling early ranking decisions at a speed and scale that traditional physics-based simulation workflows often cannot match for large screening campaigns.

Important note on interpretation: IC50 is assay-dependent, and ΔG is formally linked to equilibrium constants (Kd or Ki) under defined assumptions. Treat predicted IC50 and ΔG as ranking signals, not as assay-equivalent measurements. Always validate with chemistry-aware review and experimental follow-up.

How to use Boltz-2 on Vici.bio

A. Add at least one molecule in the Molecules panel (protein, optional DNA/RNA, ligands, ions, cofactors).

B. Decide whether you need affinity outputs. Turn Affinity on for protein–ligand ranking.

C. Choose speed vs quality using Diffusion steps, Recycles, and Samples.

Run Boltz-2 on Solo or build pipelines in Workflows . Boltz-2 can co-model proteins, nucleic acids, ligands, ions, and explicit modifications in one run. Enable affinity when you want binder triage signals.

Key configuration parameters

Recycles

Refinement passes that improve convergence.

  • Use 3 to 5 for routine runs.
  • Use 5 to 10 for complex heteromers, flexible targets, or difficult interfaces.
  • Higher values increase compute cost.

Diffusion steps

Sampling depth for generation.

  • ~200 for fast iteration and early triage.
  • 400 to 600+ for higher resolution and final candidate selection.
  • Higher steps increase runtime.

Affinity

Turns on thermodynamic-style outputs.

  • True: generates binder probability, predicted potency, and energy terms.
  • False: structure-only confidence outputs.

Samples

Independent predictions to explore pose diversity.

  • 1 to 5 for initial screens.
  • 5 to 20 for flexible systems, multiple plausible poses, or hard targets.
  • More samples improve exploration but increase cost.

Molecule inputs

  • Protein: standard FASTA sequence. If designing a macrocycle, enable the Cyclic toggle to enforce N–C covalent closure.
  • DNA/RNA: provide strings. Boltz-2 can co-fold nucleic acids with proteins for motif binding analysis.
  • Ligands, ions, cofactors: add via SMILES or CCD identifiers. Include structural ions when biologically required.
  • Glycans and modifications: supply CCD codes so modifications remain explicit in the predicted complex.

Templates and restraints

  • Template (optional): bias prediction toward a known scaffold, maintain conserved folds, or stabilize difficult targets.
  • Restraints (optional): encode partial prior knowledge, such as pocket contacts, distance constraints, or covalent attachments for macrocycles and covalent inhibitors.
  • Practical tip: restraints are most valuable when the binding site is known but the model explores incorrect poses.
  • Always sanity-check: chain IDs, stoichiometry, and obvious clashes before trusting affinity readouts.

Use cases

Small molecule lead optimization

Structure + affinity signals early in the workflow.

  • Enable Affinity and rank by ΔG or pIC50.
  • Increase Samples to explore pose diversity before wet lab.

Antibody screening

Compare CDR variants with interface confidence.

  • Track local confidence (pLDDT) and interface confidence (ipTM).
  • Use higher Samples for flexible loops, especially H3.

Macrocyclic peptide design

Bind shallow grooves at PPI interfaces.

  • Compare Cyclic on vs off to test conformational locking.
  • Prioritize models with good interface confidence and plausible chemistry.

DNA motif specificity testing

Target vs off-target motif discrimination.

  • Strong target confidence plus weaker off-target confidence supports specificity hypotheses.
  • Confirm correct stoichiometry and binding mode with a quick visual review.

Mini case study

Compound A

GO
Binder probability 0.75
IC5033.6 nM
ΔG−12.0 kcal/mol
pLDDT0.96

High probability, tight predicted potency, and strong local confidence. Send to refinement and wet lab.

Compound B

KEEP
Binder probability 0.62
IC50180 nM
ΔG−10.8 kcal/mol
pLDDT0.88

Promising, but increase Samples and Diffusion steps to test pose stability around the pocket.

Compound C

DROP
Binder probability 0.48
IC50520 nM
ΔG−10.1 kcal/mol
pLDDT0.74

Borderline probability with modest pocket confidence. If chemistry and contacts do not look right, deprioritize.

Other metrics to consider

  • pLDDT: local confidence. High is stable; low often indicates flexible loops or uncertain pockets.
  • ipTM: interface confidence for complex placement. Low suggests wrong binding mode or stoichiometry.
  • Binder probability: 0 to 1 binder vs decoy-like signal. Strong IC50 with low binder probability is a red flag.
  • ΔG (kcal/mol): stability signal. More negative is typically more favorable, best for ranking within consistent context.
  • pIC50: log potency metric: pIC50 = −log10(IC50 in molar). Useful for plotting and linear comparisons.
  • Most reliable when: good MSA coverage, structured binding site, and drug-like ligand chemistry.

Most reliable when

  • You have a strong evolutionary reference (good MSA coverage) for the protein.
  • The binding site sits in a stable structured region rather than a highly flexible loop.
  • The ligand is well-posed (reasonable size, chemistry, and protonation assumptions).

Use caution and add checks

  • Disordered regions and flexible loops: low pLDDT may be correct and can weaken affinity interpretability.
  • Collapsed pockets: check missing ions/cofactors or consider a template to stabilize geometry.
  • Conflicting metrics: high confidence but low binder probability can indicate a hallucinated interface.
  • Clashes or chain labeling errors: misassigned chains or obvious clashes can distort interface confidence.
  • Multimers: incorrect stoichiometry can distort ipTM and placement. Confirm the correct number of chains.

How it works

Boltz-2 is designed to bridge strong structure prediction with affinity-oriented outputs for protein–ligand interactions. Predicted ΔG provides a thermodynamic lens on binding favorability. Because potency changes exponentially with free energy, small improvements in ΔG can translate into large shifts in equilibrium potency under consistent assumptions.

The clean relationship is between ΔG° and an equilibrium constant like Kd (or Ki). IC50 is assay-dependent and may differ from Kd/Ki depending on setup. The tool below maps ΔG to an approximate equilibrium potency to build intuition.

ΔG° = RT ln(Kd)
Kd = exp( ΔG°RT )

Play with ΔG

Adjust ΔG and temperature to see how equilibrium potency shifts.

-10.0 298 K
Equilibrium approximation derived from ΔG° = RT ln(Kd).
N/A
N/A N/A

IC50 depends on assay format, substrate concentration, binding mechanism, and kinetics. Kd/Ki are equilibrium constants under defined assumptions. Treat predicted IC50 and ΔG as ranking signals, then confirm with chemistry-aware review and experiments.

Frequently asked Questions

References

  1. Cortés, A., Cascante, M., Cárdenas, M. L. and Cornish-Bowden, A. (2001) ‘Relationships between inhibition constants, inhibitor concentrations for 50% inhibition and types of inhibition: new ways of analysing data’, The Biochemical Journal, 357(Pt 1), pp. 263–268. doi: 10.1042/0264-6021:3570263.
  2. Abramson, J. et al. (2024) ‘Accurate structure prediction of biomolecular interactions with AlphaFold 3’, Nature, 630(8016), pp. 493–500. doi: 10.1038/s41586-024-07487-w.
  3. Passaro, S. et al. (2025) ‘Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction’, bioRxiv. doi: 10.1101/2025.06.14.659707.