Appearance
Boltz-2 Local Structure & Binding
Boltz-2 is a local structure prediction and binding-oriented model. In Liatir it is used for protein structure prediction and optional protein-ligand affinity workflows.
What it does
Boltz-2 takes protein sequence input and predicts a structure file. When ligand inputs and the selected options support it, it can also produce binding-related outputs.
When to use it
Use Boltz-2 when you want a local protein structure prediction workflow that can feed directly into Liatir structure viewers and reports.
Inputs in Liatir
- Protein FASTA/FAA file, or an inline protein sequence.
- Optional ligand SMILES or CCD identifier.
- Prediction options such as output format, recycling steps, diffusion samples, MSA server use, and accelerator choice.
Outputs
Liatir can produce:
- PDB or mmCIF structure files when prediction succeeds;
- confidence and binding metadata where available;
- logs and runtime warnings;
- viewer-compatible artifacts;
- provenance with model, runtime, input, parameters, and output files.
Hardware and installation
Boltz-2 can be installed for CPU-only use, but CPU inference can take a long time. CUDA GPU inference is strongly preferred for serious use.
Liatir installs Boltz-2 in an isolated managed Python runtime. The runtime currently requires Python 3.10, 3.11, or 3.12.
Limits and cautions
Structure prediction can be slow and resource-heavy. A completed run without a structure file is treated as an error in Liatir because the main scientific artifact is missing.
Predicted structures and affinities should be interpreted carefully and, where important, validated with additional methods.
The structure file is the main artifact. If a run produces logs but no PDB or mmCIF file, the scientific output is incomplete.