Boltz-1
Fully Open-Source Biomolecular Complex Foundation Model by MIT & Valence
State-of-the-art open alternative to AlphaFold 3 under MIT license for proteins, RNA, DNA, and chemical ligands.
Technical Overview & Biological Significance
Boltz-1 is the first fully open-source foundation model achieving AlphaFold 3-level accuracy in predicting 3D structures of complex biomolecules including proteins, RNA, DNA, and covalent modifications. Released by Valence Labs and MIT researchers under the permissive MIT license, Boltz-1 provides both training code, model weights, and fast local inference pipelines for academic labs and biotech enterprises alike.
The Biological Challenge Solved
While proprietary and semi-open models advance structural biology, their non-commercial terms and cloud restrictions prevent enterprise deployment, high-throughput virtual screening, and integration into proprietary drug discovery pipelines. Boltz-1 democratizes frontier biomolecular modeling without commercial restrictions.
Key Computational Capabilities
Full MIT Open Source Weights
Zero commercial restrictions; deployable on local GPU clusters, on-premise servers, and cloud providers.
AlphaFold 3 Competitive Accuracy
Benchmarks on recent PDB complexes show competitive accuracy with AlphaFold 3 on protein-protein and protein-ligand interfaces.
Command-Line & Python API
Simple YAML-based input definitions and standard CLI commands for automated high-throughput batch execution.
Integrated MSA Generation
Supports remote MMseqs2 MSA servers or local databases for high-efficiency alignment generation.
Executable Commands & API Usage
Install boltz via pip and predict complex structure from a simple YAML specification:
# 1. Install boltz package
pip install boltz
# 2. Run prediction using the public MSA server
boltz predict input_complex.yaml --use_msa_server --out_dir ./boltz_results
# 3. Inspect generated mmCIF and confidence scores
cat ./boltz_results/confidence.json | jq .Input Specifications & Inference Results
version: 1
sequences:
- protein:
id: A
sequence: MKTIIALSYIFCLVFA
- dna:
id: B
sequence: ATGCCGTAGCTADefine IDs, types (protein, dna, rna, ligand), and sequence strings in a human-readable YAML document.
{
"confidence_score": 0.88,
"ptm": 0.90,
"iptm": 0.85,
"complex_plddt": 87.4
}Calibrated quantitative scores ready for downstream analysis.
Biological Result Interpretation Guide
How researchers interpret confidence thresholds, fold-change values, and functional impact:
Can be embedded directly into pharmaceutical virtual screening funnels and proprietary machine learning workflows.
Closely tracks AlphaFold 3 interface metrics, enabling direct comparison of predictions between models.
Boltz-1 vs. Traditional Computational Methods
| Dimension | Boltz-1 (This Tool) | Traditional Pipelines |
|---|---|---|
| License | Permissive MIT (Commercial friendly) | Proprietary or Non-commercial restricted |
| Local Deployment | Full offline local execution | Cloud-only web server |
| Weights Availability | Freely downloadable on Hugging Face | Closed or gated weights |
Frequently Asked Questions (FAQ)
Is Boltz-1 completely free to use in commercial projects?
Yes, Boltz-1 is released under the MIT license, allowing unrestricted academic, open-source, and commercial biotechnology applications.
What hardware do I need to run Boltz-1 locally?
An NVIDIA GPU with at least 16GB VRAM (such as an RTX 4080/4090 or A10G) is recommended for proteins up to 800 residues.