google-deepmind/alphafold3
AlphaFold 3 inference pipeline.
What it solves
AlphaFold 3 predicts the accurate 3D structures of biomolecular interactions, allowing researchers to model how proteins, nucleic acids, and other molecules interact with one another.
How it works
It uses an inference pipeline that processes input sequences (provided via JSON) through a data pipeline for genetic and template searches (CPU-based) and an inference engine (GPU-based) to predict the final molecular structure. The system can be deployed via Docker and requires specific model parameters provided by Google.
Who it’s for
It is designed for scientists and researchers in biology and chemistry who need to predict biomolecular structures for theoretical modeling.
Highlights
- Supports the prediction of complex biomolecular interactions.
- Includes specialized model versions, such as one for watermarked structures (SynthID Bio-structure) and one fine-tuned for leaving atom handling (AF3-LA).
- Provides a clear separation between the CPU-intensive data pipeline and the GPU-intensive inference process.
- Integrated with various scientific libraries like JAX, Haiku, and RDKit.
Related
- Project
- Project
- Project
- Dispatch
- Project