MIT and Recursion have released Boltz-2, the first open-source biomolecular co-folding model that jointly predicts 3D complex structures and molecular binding affinity with unprecedented speed and precision.
The AI model approaches the accuracy of free energy perturbation (FEP), a gold-standard physics-based method, while performing up to 1,000 times faster, significantly reducing cost and time for large-scale molecular screening.
Boltz-2 outperformed all CASP16 affinity challenge participants and is released under the permissive MIT license, making the model, weights, and training pipeline available for both academic and commercial use.
The model was trained using Recursion's NVIDIA-powered supercomputer BioHive-2 and incorporates molecular dynamics simulations and approximately 5 million binding affinity measurements.