Boltz-2: Structure Prediction

Biomolecular structure prediction for proteins, nucleic acids, small molecules, and their complexes. Boltz-2 leverages advanced deep learning diffusion models to generate highly accurate 3D structures from sequence information, providing confidence metrics through pLDDT and pTM scores. This state-of-the-art tool enables researchers to predict protein structures, antibody-antigen complexes, and multi-chain assemblies with unprecedented speed and reliability for drug discovery applications.

GitHub: https://github.com/jwohlwend/boltz

Key Features

  • Protein Structure Prediction: From sequence to 3D structure
  • Confidence Scoring: pLDDT and pTM metrics for quality assessment
  • Multi-chain Support: Antibody heavy/light chains, protein complexes
  • Fast Inference: Suitable for high-throughput applications

Input & Configuration

Sequence Input

  • FASTA format: Protein or nucleic acid sequences
  • Multi-chain support: Heavy/light chains for antibodies
  • Length range: 20-2500 residues (optimal: 50-800)

Key Parameters

  • Output Format: PDB
  • Recycling Steps: 3-5 iterations
  • Sampling Steps: 200 (diffusion denoising)
  • Diffusion Samples: 1-10 for prediction diversity

Confidence Assessment

Quality Metrics

  • pLDDT > 0.9: Very high confidence, near-atomic quality
  • pLDDT 0.7-0.9: Confident regions, reliable for applications
  • pLDDT 0.5-0.7: Low confidence, interpret with caution
  • pLDDT < 0.5: Very low confidence, likely disordered

Boltz-2 Confidence Score: Calculated as 0.8 × overall pLDDT + 0.2 × interface pTM

Integration Workflows

  • Structure prediction → DiffDock-PP → Protein complex analysis
  • Sequence input → Boltz-2 → AutoDock Vina screening
  • Boltz-2 Report - Structure prediction analysis and quality assessment
  • Mol* - Structure visualization

Tutorials