MIT scientists have introduced Boltz-1, an innovative open-source AI model poised to revolutionize biomedical research and drug development.
Created by a talented group at the MIT Jameel Clinic for Machine Learning in Health, Boltz-1 stands as the first fully open-source model to reach state-of-the-art performance comparable to AlphaFold3, developed by Google DeepMind, which excels at predicting the 3D structures of proteins and other biological molecules.
The development of Boltz-1 was led by MIT graduate students Jeremy Wohlwend and Gabriele Corso, alongside MIT Jameel Clinic Research Affiliate Saro Passaro and prominent professors Regina Barzilay and Tommi Jaakkola from the School of Electrical Engineering and Computer Science. Wohlwend and Corso showcased the model at an event on December 5 at MIT’s Stata Center, emphasizing their vision for global collaboration and the acceleration of scientific discoveries in biomolecular modeling.
“We envision this as just the beginning for the community,” Corso remarked. “The name Boltz-1 signifies that this is not the final product. We eagerly seek contributions from the scientific community.”
Proteins are integral to nearly all biological processes, and their functionality is closely tied to their structure. Therefore, accurately understanding protein structures is vital for developing new drugs or engineering proteins with desired characteristics. However, predicting the complex way in which a protein’s amino acid chain folds into its 3D shape has presented significant hurdles for researchers over the years.
DeepMind’s AlphaFold2, which earned its creators Demis Hassabis and John Jumper the Nobel Prize in Chemistry in 2024, uses advanced machine learning techniques to predict 3D protein structures at an extraordinary accuracy level, rivaling those determined experimentally. This model has significantly contributed to advancements in drug development across academic and commercial laboratories globally.
AlphaFold3 enhances its predecessors by integrating a generative AI model, specifically a diffusion model, which effectively addresses the uncertainties in predicting intricate protein structures. In contrast to AlphaFold2, which operates as an open-source model, AlphaFold3 is not entirely open for public or commercial use, leading to scrutiny from the scientific community and sparking a global demand for a commercially usable version.
MIT’s team approached their work on Boltz-1 by initially replicating AlphaFold3’s methodology but then innovating upon the diffusion model and incorporating enhancements that maximized the model’s accuracy, including algorithms designed to increase prediction efficiency.
Beyond just the model, the MIT team has shared their entire training and fine-tuning pipeline as open-source, inviting other researchers to build on Boltz-1’s foundation.
“I’m incredibly proud of Jeremy, Gabriele, Saro, and the entire Jameel Clinic team for bringing this project to fruition. Their dedication and effort over countless days and nights have culminated in this achievement. We’re excited about potential future enhancements and look forward to sharing more in the coming months,” stated Barzilay.
It took the MIT team four intense months of experimentation to develop Boltz-1. One of the primary challenges was navigating the ambiguity and diversity of data within the Protein Data Bank—a comprehensive collection of biomolecular structures explored by scientists over the past 70 years.
“I spent many late nights grappling with this wealth of data. Much of it requires domain-specific knowledge—there really are no shortcuts,” noted Wohlwend.
In their findings, the MIT researchers demonstrated that Boltz-1 achieved a level of accuracy on par with AlphaFold3 when predicting a diverse array of complex biomolecular structures.
“What Jeremy, Gabriele, and Saro have achieved is truly impressive. Their diligence has made biomolecular structure prediction far more accessible, paving the way for significant breakthroughs in molecular sciences,” said Jaakkola.
The research team intends to further enhance Boltz-1’s performance and reduce the time required for predictions. They also encourage fellow scientists to test Boltz-1 via their GitHub repository and connect with the Boltz-1 user community through their Slack channel.
“We believe there’s still ample opportunity for advancements in these models. We’re eager to collaborate and see how the community leverages this tool,” reiterated Wohlwend.
Mathai Mammen, CEO and president of Parabilis Medicines, hailed Boltz-1 as a “groundbreaking” model. “By making this model accessible, the MIT Jameel Clinic team is opening doors to advanced structural biology tools. This monumental effort will expedite the development of transformative medicines. Kudos to the Boltz-1 team for achieving this significant milestone!”
“Boltz-1 will greatly empower my lab and the larger community,” added Jonathan Weissman, an MIT biology professor affiliated with the Whitehead Institute for Biomedical Engineering, who did not participate in the study. “I expect a significant wave of discoveries to emerge from democratizing this potent tool.” Weissman further anticipates that the open-source nature of Boltz-1 will inspire a myriad of innovative applications.
This significant research initiative was supported by a U.S. National Science Foundation Expeditions grant, the Jameel Clinic, the U.S. Defense Threat Reduction Agency, and the MATCHMAKERS project funded by Cancer Research UK and the U.S. National Cancer Institute.
Photo credit & article inspired by: Massachusetts Institute of Technology
