Structure Prediction
Introduction
Boltz-1 is a machine learning tool that performs at the level of AlphaFold3 for biomolecular interactions without any restrictions on commercial use, or input types. There were promising implementations by the community beforehand, including from the teams at Ligo Biosciences, Chai Discovery, Baidu and ByteDance, but these either had restrictions, or were incomplete reproductions.
Performance
The team evaluated model performance on CASP15 and a custom PDB set curated for benchmarking all-atom structure predictions. The custom set excludes chains whose sequences are found in the training set. It is also constructed to optimize for diversity by iteratively adding structures belonging to new unseen clusters.
Boltz-1 matches Chai-1, and therefore Alphafold3, in performance across LDDT, TM, DockQ, lDDT-PLI, and Ligand RMSD.
Architecture
The model adds a novel Dense MSA Pairing Algorithm that optimally balances sequence signal quality with computational efficiency, alongside a Unified Cropping Algorithm that combines spatial and contiguous cropping strategies. The model also implements a more flexible pocket-conditioning approach that eliminates the need for separate models and better handles partial binding site information.
In addition accuracy improvements, the model features improvements in performance and training efficiency. These include a strategic reordering of MSA module operations to better integrate different types of representations, modified transformer layers with proper residual connections, and the introduction of Kabsch diffusion interpolation to address theoretical limitations in the original AlphaFold3 approach. The confidence model was also completely redesigned to leverage pretrained trunk weights and incorporate features from the diffusion model, leading to more robust predictions.
Applications
Boltz-1 offers great potential for a wide range of tasks including antibody structure prediction, protein-ligand cofolding for enzyme discovery, protein protein interaction predictions. This also includes DNA and RNA sequences as input. It address many of the limitations of Alphafold, and paves the way for open source collaboration in the field.
Confidential, Secure, Scalable Interface
At Tamarind, we offer hundreds of the leading protein design, structure, and sequence based modeling tools with up to hundreds of thousands of sequences as input.
Try Boltz-1 on Tamarind and get in touch to learn more on our secure, confidential cloud offering. We also support a programmatic interface, Boltz-1 finetuning, and running tools at very high throughput. You own all of your data, inputs and derivatives!
Dozens of biotechs and pharmas including Adimab, Septerna, and Allozymes use Tamarind to accelerate their drug discovery, protein design, and enzyme engineering work.