BREAKING: Boltz-2 Now Live On Rowan
(Apologies for the repeat email today!)
This morning, a team of researchers from MIT and Recursion released Boltz-2, an open-source protein–ligand co-folding model. We listed the predecessor model Boltz-1 alongside Chai-1 earlier this month, writing that this field was “one of the most exciting areas in computer-assisted drug design right now,” and this new release only underscores that point.
Just like Boltz-1, Boltz-2 can predict the 3D structure of biomolecular complexes from sequence data and SMILES strings, showing improved accuracy across a range of benchmarks (paper). But Boltz-2 can also predict protein–ligand binding affinity. The authors show that Boltz-2 performs almost as well as the industry-standard FEP+ workflow and handily outperforms cheaper physics-based methods like MM/PBSA, although performance is considerably worse on internal targets from Recursion:
Full assessment of Boltz-2’s capabilities will require extensive benchmarking and external verification; in particular, we’re curious how Boltz-2 will compare to other protein–ligand binding-affinity models like PLAPT or GNINA. But while we await these studies, it’s already possible to start using Boltz-2 for binding-affinity prediction through Rowan.
Rowan’s protein–ligand co-folding workflow is available for all users; simply select “Boltz-2” as your co-folding model, input your desired protein and ligand, and submit the calculation. The calculation will run on cloud GPU hardware and predict a protein–ligand complex and IC50 value within a few minutes.
We wanted to get this out to our users as fast as possible, and are excited to hear how this model works in the wild. If you have any feedback, please reach out at contact@rowansci.com!