Boltz

Boltz predicts the 3D structures of biomolecular complexes and binding affinity using deep learning models.

Key Features:

• Structure prediction: Boltz-1 achieves AlphaFold 3-level accuracy in predicting the 3D structures of biomolecular complexes.
• Boltz-steering: An inference time steering technique that fixes hallucinations and non-physical predictions from the models.
• Binding affinity prediction: Boltz-2 approaches the performance of free-energy perturbation methods in estimating small molecule-protein binding affinity.
• Controllability features: Boltz-2 includes experimental method conditioning, distance constraints, and multi-chain template integration for structure prediction.
• Computational efficiency: Boltz-2 is at least 1000 times more computationally efficient than free-energy perturbation methods.
• Open-source release: Training and inference code, model weights, datasets, and benchmarks are released under the MIT open license.

Scientific Applications:

• Drug discovery: Boltz-2 coupled with a generative model for small molecules enables finding diverse, synthesizable, high-affinity binders.
• Protein design: Understanding biomolecular interactions supports advancing protein design.
• Structural biology: Predicting biomolecular complex structures provides a foundation for academic and industrial research.
• Binding affinity estimation: Estimating small molecule-protein binding affinity with strong correlation to experimental readouts on many benchmarks.

Methodology:

Boltz-1 incorporates innovations in model architecture, speed optimization, and data processing. Boltz-2 is a structural biology foundation model. Boltz-steering is an inference time steering technique.