sokrypton/ColabFold

What it solves

ColabFold® is designed to make high-accuracy protein folding predictions accessible to everyone by providing easy-to-use Google Colab notebooks and a command-line interface. It removes the technical barriers to running complex models like AlphaFold2, AlphaFold3, and ESMFold, which typically require massive databases and significant computational resources.

How it works

The project provides a suite of notebooks and tools that integrate several protein structure prediction models. It leverages a public MSA (Multiple Sequence Alignment) server to generate the necessary sequence alignments quickly, avoiding the need for users to download and store nearly a terabyte of genomic databases locally. For local installations, it provides colabfold_batch and colabfold_search for structure prediction and MSA generation, with support for GPU acceleration via MMseqs2-GPU.

Who it’s for

It is primarily for biologists, biochemists, and researchers in structural biology who need to predict the 3D structures of proteins (monomers and complexes) without needing to manage complex bioinformatics pipelines or high-performance computing clusters.

Highlights

• Multi-model support: Integrates AlphaFold2, AlphaFold3 (OpenFold3), ESMFold, RoseTTAFold2, and Boltz.
• Cloud-accessible: Provides Google Colab notebooks for immediate use without installation.
• Efficient MSA generation: Uses a public server to handle the sequence search, significantly reducing local storage requirements.
• Local execution: Supports installation via conda, pip, or Docker for local GPU-accelerated predictions.
• Advanced input: Supports non-protein molecules like DNA, RNA, and ligands (via SMILES or CCD codes) for AlphaFold3-compatible inputs.
• Performance optimizations: Includes JAX compilation caching and Pallas/Triton kernels for faster predictions on Ampere GPUs.