MIT and Recursion partner for AI drug discovery tool

A team led by Regina Barzilay, a computer science professor at the Massachusetts Institute of Technology, has launched Boltz-2, an algorithm that unites protein folding and prediction of small-molecule binding affinity in one package. The artificial intelligence model, which was developed with insights from the machine learning–driven company Recursion Pharmaceuticals, is openly available for academic and commercial use.

The field of AI-powered prediction of protein structure has developed rapidly in recent years, but for drug hunters, using AI to find small molecules that bind to a particular target has remained a challenge. It’s one that many academic groups and private companies want to solve.

“Affinity was an open problem for decades,” Barzilay says. “It really required very novel machine learning to develop this technique.”

The researchers say their new AI model approaches the level of accuracy achieved by traditional computational chemistry—such as methods involving free-energy perturbation calculations—but much more quickly and cheaply. The model is described in a preprint and has not been peer-reviewed.

Boltz-2 was trained on a variety of data sources, including internal data from Recursion. The firm, which also provided computational support, had been using Boltz-1. But Recursion says it is getting better results since replacing that earlier version with Boltz-2, not just for virtual screening and hit identification but also for lead optimization.

Recursion’s chief R&D officer, Najat Kahn, emphasizes that experimentation will still be needed. “You can actually improve the algorithms as you learn from the experimental validation as well,” she says.