Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction.

Area of Science:

• Structural biology
• Computational biology
• Artificial intelligence in life sciences

Background:

• Accurate modeling of biomolecular interactions is crucial for understanding molecular function and drug development.
• Existing models like AlphaFold3 and Boltz-1 excel at structure prediction but struggle with binding affinity.
• Predicting binding affinity is essential for assessing molecular function and therapeutic potential.

Purpose of the Study:

• Introduce Boltz-2, a foundation model for enhanced biomolecular structure and binding affinity prediction.
• Develop an AI model that rivals traditional methods in accuracy while significantly improving computational efficiency.
• Provide a controllable and extensible AI framework for advancing drug discovery.

Main Methods:

• Developed Boltz-2, a structural biology foundation model with advanced controllability features.
• Integrated experimental method conditioning, distance constraints, and multi-chain template integration for structure prediction.
• Evaluated Boltz-2's performance against established methods like free-energy perturbation (FEP) for binding affinity prediction.

Main Results:

• Boltz-2 demonstrates strong performance in both biomolecular structure and binding affinity prediction.
• The model shows a strong correlation with experimental data across multiple benchmarks.
• Boltz-2 achieves binding affinity prediction performance comparable to FEP methods but is over 1000x more computationally efficient.

Conclusions:

• Boltz-2 offers a significant advancement in predicting biomolecular interactions, addressing limitations in current AI models.
• The model's efficiency and accuracy present a powerful tool for accelerating drug discovery workflows.
• The open release of Boltz-2's code and weights promotes further research and innovation in machine learning for biology.