aqlaboratory/openfold-3

What it solves

OpenFold3-preview is an open-source reproduction of DeepMind's AlphaFold3, designed to provide a fully open and maintainable biomolecular structure prediction model. It allows researchers to predict the 3D structures of proteins, RNA, DNA, and small molecules, overcoming the limitations of closed-source models.

How it works

The project replicates the input features and architecture of AlphaFold3. It uses pipelines for generating Multiple Sequence Alignments (MSAs) via the ColabFold server or tools like JackHMMER and hhblits. To improve performance, it utilizes kernel acceleration through cuEquivariance and DeepSpeed4Science kernels.

Who it’s for

It is intended for biologists, biochemists, and AI researchers who need high-accuracy biomolecular structure prediction for academic or commercial use under an Apache 2.0 license.

Highlights

• Broad Modality Support: Predicts structures for standard and non-canonical protein, RNA, and DNA chains, and small molecules.
• Kernel Acceleration: Uses cuEquivariance and DeepSpeed4Science for efficient, high-speed inference.
• Scalability: Supports multi-query jobs and distributed predictions across multiple GPUs.
• Low-Memory Mode: Includes custom settings for users with memory-constrained GPU resources.
• Open Training Data: Provides full training data, including a reproduction of the MGnify-based 13M-sequence distillation dataset.
• Training Capabilities: Includes tools for data processing and model training.