MIT and Recursion have released Boltz-2, a next-generation artificial intelligence model designed to predict the binding affinity between molecules and protein targets with what the companies describe as “unprecedented” speed, scale, and accuracy.
The companies said that Boltz-2 is trained on proprietary high-resolution biochemical data generated at Recursion and uses a modified large language model architecture to outperform existing tools. Recursion claims Boltz-2 achieves a 30% improvement in root mean square error (RMSE) compared to state-of-the-art methods, including AlphaFold 2.3 and RoseTTAFold All-Atom.
“We believe Boltz-2 is a major leap forward,” said Ben Mabey, CTO at Recursion, adding that it allows researchers to “model the physical world of chemistry with large-scale generative AI.”
Boltz-2 was developed in collaboration with the Green Lab at MIT, where scientists used DiffDock, a molecular docking model, to help scale training to over 20 billion predictions. The company said this effort is one of the largest model training datasets of its kind.
Recursion stated that Boltz-2 is now integrated into its internal drug discovery pipeline and may be released publicly to advance open science.
The company emphasized its goal to bridge machine learning with wet lab experimentation to accelerate the design and development of new therapeutics.
