DeepMind’s AlphaFold2 uses machine learning to predict protein structures quickly and accurately, matching results from lab experiments. This model has been used widely in research, helping advance drug development.
AlphaFold3 improves on AlphaFold2 by using a new AI technique called a diffusion model, which handles the uncertainty of predicting complex protein structures. However, unlike AlphaFold2, AlphaFold3 is not fully open source or available for commercial use, leading to criticism from scientists and sparking a global race to create a commercial version.
MIT scientists have unveiled Boltz-1, a powerful open-source AI model that could speed up biomedical research and drug development. Created by the MIT Jameel Clinic for Machine Learning in Health, Boltz-1 is the first fully open-source model to achieve top-tier performance, similar to AlphaFold3 from Google DeepMind, which predicts the 3D structures of proteins and other biological molecules.
Gabriele Corso from MIT said, “We hope for this to be a starting point for the community. There is a reason we call it Boltz-1 and not Boltz. This is not the end of the line. We want as much contribution from the community as we can get.”
The MIT researchers behind Boltz-1 initially used the same approach as AlphaFold3 but then explored ways to improve the underlying diffusion model. They focused on incorporating enhancements that significantly boosted the model’s accuracy, such as new algorithms to improve prediction efficiency.
In addition to releasing the model, they also open-sourced their entire pipeline for training and fine-tuning, allowing other scientists to build upon and further develop Boltz-1.
The MIT team spent four months and conducted numerous experiments to develop Boltz-1. One of their biggest challenges was dealing with the ambiguity and variety in the Protein Data Bank, which contains biomolecular structures solved by biologists over the past 70 years.
After overcoming these hurdles, their experiments showed that Boltz-1 achieves the same level of accuracy as AlphaFold3 in predicting a wide range of complex biomolecular structures.
The researchers plan to keep improving Boltz-1’s performance and reduce the time it takes to make predictions. They also encourage other scientists to try out Boltz-1 on their GitHub repository and join the community of users on their Slack channel for collaboration and support.
Journal Reference:
- Jeremy Wohlwend, Gabriele Corso, Saro Passaro et al. Boltz-1: Democratizing Biomolecular Interaction Modeling. Paper
