Boltz-2: A New AI Model for Structure and Binding Affinity Prediction - a blog post by ESR5 Simone Mariani

​Researchers from MIT, Valence Labs, Recursion, and ETH Zurich have developed Boltz-2, a foundation model that advances both biomolecular structure prediction and binding affinity estimation. This work addresses a longstanding computational challenge in drug discovery: accurately predicting how tightly small molecules bind to protein targets.
 
Current methods for binding affinity prediction face significant limitations. Free energy perturbation (FEP) simulations provide high accuracy but require substantial computational resources, often taking days to evaluate small compound sets. Faster approaches like molecular docking can screen large libraries quickly but lack the precision needed for reliable drug development decisions. Boltz-2 attempts to bridge this performance-speed gap.
 
The model builds on previous structure prediction advances like AlphaFold3 and Boltz-1, incorporating several key innovations. The training dataset combines experimental structures from the Protein Data Bank with molecular dynamics ensembles, exposing the model to both static equilibrium states and dynamic fluctuations. The researchers curated millions of binding affinity measurements from public databases, standardizing diverse experimental protocols and filtering for data quality. The architecture includes specialized components for both structure prediction and affinity estimation, with the affinity module operating on the model's structural representations.
 
On the FEP+ benchmark, Boltz-2 achieved correlation coefficients approaching those of FEP methods while running over 1000 times faster. In the CASP16 affinity challenge, it outperformed all submitted entries without specialized tuning. The model also demonstrated practical utility in virtual screening experiments, successfully identifying high-affinity binders for the TYK2 kinase target when validated against absolute binding free energy calculations.
 
The researchers acknowledge several limitations, including variable performance across different protein families and challenges with large conformational changes upon binding. They note that accurate structure prediction remains a prerequisite for reliable affinity estimation.
 
Boltz-2's code, model weights, and training data are being released under an open license, providing the scientific community with access to both the trained model and the complete training pipeline for further development and application.