Drug discovery laboratories may be able to innovate more rapidly with a newly available AI model.
Dubbed Boltz-2, the new model is the successor to Boltz-1, a deep learning model like AlphaFold that predicts biomolecular interactions. Boltz-2’s major new offering is a binding affinity prediction model that achieves results comparable to free-energy perturbation (FEP) simulations but runs 1,000 times faster.
Advanced Lab Management Certificate
The Advanced Lab Management certificate is more than training—it’s a professional advantage.
Gain critical skills and IACET-approved CEUs that make a measurable difference.
Like its predecessor, Boltz-2 is under an open-source license from the Massachusetts Institute of Technology (MIT) and has no restrictions on academic or commercial use. According to a press release, Boltz-2 joint modeling of complex structures and binding affinities makes “accurate in silico screening practical for early-stage drug discovery.”
New model promises faster and more efficient research
An important component of drug design is assessing how strongly a drug molecule will bind to its target protein, which is referred to as its binding affinity. Typically, binding affinity is predicted with a physics-based technique, FEP. While accurate, FEP can be time- and resource-intensive, limiting research progress.
Boltz-2’s binding affinity prediction engine can run 1000 times faster than FEP while achieving comparable accuracy, representing an opportunity for biopharma labs.
“Boltz-2’s structure and affinity prediction allow a very cheap but relatively accurate way of testing compounds before they are ordered,” says Gabriele Corso, a graduate student at MIT and one of the lead developers of the project. “Moreover, [by] fully leveraging virtual screening pipelines built on top of Boltz-2, scientists can reduce the number [of] experimental validation[s] needed to find effective molecules.”
When asked what the key was to developing the new binding affinity model, Corso explains:
“Models like AlphaFold and Boltz have developed [a] strong understanding of physical interactions without the need of simulations. Boltz-2 is able to exploit this understanding by having an extra module trained on millions of training datapoints from publicly available biochemical assays. This allows [for] accurate binding affinity calculations without the need for extensive simulations.”
How labs can implement Boltz-2
The technical requirements to run Boltz-2 have remained the same as those of Boltz-1. Researchers will need access to a computer—either locally or through the cloud—outfitted with modern GPUs. Forty gigabytes of video RAM (VRAM) in the GPUs is ideal, but even 32 or 24 gigabytes should suffice for most input sizes. Corso notes that a collaboration with NVIDIA enabled the Boltz developers to reduce the runtime and memory footprint for the model. As a result, Boltz-2 can predict the structure of larger complexes better than Boltz-1, even with the same amount of VRAM.
Interested lab managers can find the source code and detailed installation instructions on the Boltz GitHub repository.
For those interested in keeping up with the development of Boltz and connecting with other users, the Boltz team has a public Slack channel available to join.
