Scientists from MIT have unveiled Boltz-1, a cutting-edge, open-source AI model designed to significantly boost progress in biomedical research and drug development.
Created by researchers at the MIT Jameel Clinic for Machine Learning in Health, Boltz-1 stands out as the first fully open-source model that matches the high performance of AlphaFold3, the Google DeepMind model known for predicting the 3D structures of proteins and other biological molecules.
MIT graduate students Jeremy Wohlwend and Gabriele Corso led the development of Boltz-1, working alongside MIT Jameel Clinic Research Affiliate Saro Passaro and professors Regina Barzilay and Tommi Jaakkola from the Electrical Engineering and Computer Science department. At a presentation on December 5 at MIT’s Stata Center, Wohlwend and Corso expressed their aim to promote global collaboration, speed up discoveries, and offer a strong platform for advancing biomolecular modeling.
“We consider this a starting point for the community,” Corso explained. “We named it Boltz-1 rather than just Boltz because this isn’t the final version. We welcome as much input from the community as possible.”
Proteins are crucial for almost all biological processes. The shape of a protein is tightly linked to its function, making it essential to understand its structure for developing new drugs or engineering proteins with specific functions. However, predicting a protein’s 3D structure from its amino acid sequence has been a longstanding challenge due to the complexity of the folding process.
DeepMind’s AlphaFold2, which earned creators Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry, uses machine learning to quickly predict 3D protein structures with remarkable accuracy. This open-source model has been influential in drug development, utilized by research teams worldwide.
AlphaFold3 improves on previous versions by including a generative AI model, known as a diffusion model, which handles the uncertainty in predicting complex protein structures more effectively. However, unlike AlphaFold2, AlphaFold3 is not fully open-source and isn’t available for commercial use, leading to criticism and a global push to create a commercially available version.
The MIT team, while developing Boltz-1, began with a similar approach to AlphaFold3 but modified the underlying diffusion model to enhance accuracy. They introduced new algorithms to increase prediction efficiency and open-sourced their entire training and fine-tuning pipeline for other scientists to build upon Boltz-1.
“I’m incredibly proud of Jeremy, Gabriele, Saro, and the entire Jameel Clinic team for this achievement. This project required countless hours of dedication to reach this point. We have many exciting ideas for further enhancements and look forward to sharing them soon,” stated Barzilay.
It took the MIT team four months and numerous experiments to develop Boltz-1. One major challenge was navigating the ambiguity and diversity in the Protein Data Bank, a repository of biomolecular structures solved over the past 70 years.
“There were many long nights spent struggling with this data. It’s a lot of domain knowledge that one must acquire without shortcuts,” Wohlwend mentioned.
Their experiments demonstrated that Boltz-1 matches AlphaFold3’s accuracy in predicting a range of complex biomolecular structures.
“What Jeremy, Gabriele, and Saro have achieved is extraordinary. Their dedication has made biomolecular structure prediction more accessible, which will transform advancements in molecular sciences,” said Jaakkola.
The team plans to continue enhancing Boltz-1’s performance and reducing prediction times. They encourage researchers to explore Boltz-1 on their GitHub repository and connect with other users via their Slack channel.
“We believe there’s still much work to be done to improve these models. We’re eager to collaborate with others and see how the community utilizes this tool,” Wohlwend added.
Mathai Mammen, CEO and president of Parabilis Medicines, hailed Boltz-1 as a “breakthrough” model. “By open-sourcing this advance, the MIT Jameel Clinic and collaborators are democratizing access to state-of-the-art structural biology tools,” he said. “This landmark effort will speed up the development of life-changing medicines. Thanks to the Boltz-1 team for driving this significant leap forward!”
“Boltz-1 will be incredibly beneficial for my lab and the entire community,” added Jonathan Weissman, an MIT professor of biology and member of the Whitehead Institute for Biomedical Engineering, who was not involved in the study. “We’ll witness a wave of discoveries enabled by democratizing this powerful tool.” Weissman predicted that Boltz-1’s open-source nature would lead to a wide range of innovative applications.
This work received support from a U.S. National Science Foundation Expeditions grant; the Jameel Clinic; the U.S. Defense Threat Reduction Agency’s DOMANE program; and the MATCHMAKERS project funded by the Cancer Grand Challenges partnership, supported by Cancer Research UK and the U.S. National Cancer Institute.
