ColabFold: What It Is and How It Differs from AlphaFold2

Short answer: ColabFold is an open-source protein structure prediction tool that pairs the fast MMseqs2 homology search with AlphaFold2 (or RoseTTAFold) models, so you can run predictions from a Google Colab notebook, a command line, or a local install. It differs from DeepMind’s original AlphaFold2 notebook mainly in how it builds the multiple sequence alignment (MMseqs2 instead of jackhmmer).

This guide is for researchers and lab staff who have a protein sequence and need to know which ColabFold entry point fits, what is different from AlphaFold2, and what to expect. It is informational; it is not a protocol for any specific project.

Definition and scope

ColabFold was described by Mirdita and colleagues in Nature Methods (2022;19:679-682, “ColabFold: making protein folding accessible to all”). The abstract states that it combines the fast homology search of MMseqs2 with AlphaFold2 or RoseTTAFold, that the 40-60-fold faster search and optimized model use allow prediction of close to 1,000 structures per day on a server with one graphics processing unit, and that coupled with Google Colaboratory it becomes a free and accessible platform for protein folding. It is open source, and the project’s GitHub repository (sokrypton/ColabFold, MIT license, page listing v1.6.3 when read on 2026-11-08) hosts the notebooks and command-line tools.

It is a prediction tool for protein structures from sequence. It does not replace experimental structure determination such as X-ray crystallography.

How ColabFold is used in practice

The repository describes three ways in. The table lists each entry point and what the repository says about it.

Entry point	What the repository says
Google Colab notebooks	The AlphaFold2_mmseqs2 notebook runs in the browser. Sequence-length limits depend on the free GPU that Colab provides.
colabfold_batch	Takes a FASTA or CSV file and queries the public MSA server. Installed with pip install colabfold[alphafold,openmm]. Supports AlphaFold2-multimer for complexes.
LocalColabFold	An installer script by Yoshitaka Moriwaki for running on your own machine, including Windows through WSL2.

Two practical points come from the repository. First, the first run for a new sequence length or MSA depth includes a compile step that can take minutes; a JAX cache helps on later runs. Second, the public MSA server asks users not to query it from multiple computers. The repository also describes a --use-fast-kernels option (about 2.5 times faster, with slightly lower memory use) and GPU-accelerated MMseqs2 search with an optional GPU server.

ColabFold compared with the original AlphaFold2 notebook

Dimension	DeepMind AlphaFold2 notebook	ColabFold
MSA generation	jackhmmer	MMseqs2 (AlphaFold2_mmseqs2)
Search speed	Baseline	40-60-fold faster search (paper abstract)
Models	AlphaFold2	AlphaFold2, AlphaFold2-multimer, or RoseTTAFold (paper and repository)
Ways to run	Notebook	Notebook, command line, local install
License	See DeepMind’s repository	MIT (ColabFold repository)

For precomputed predictions you do not need to run anything: see the AlphaFold Protein Structure Database guide. For other prediction routes, see ESMFold (a language-model approach) and SWISS-MODEL (homology modelling). Predicted coordinates can be inspected in a viewer such as PyMOL, and experimental structures are archived in the Protein Data Bank.

Where it matters for lab decisions

• Compute: the notebook route depends on the GPU Colab allocates, so long sequences or complexes may not fit; command-line and local routes shift that to hardware you control.
• Shared infrastructure: batch work through the public MSA server should come from one machine, per the repository’s request.
• Data handling: sequences sent to a public server leave your environment; check your institution’s rules for unpublished or proprietary sequences before using it.

Common misunderstandings

• “ColabFold is a different model from AlphaFold.” It runs AlphaFold2 (and RoseTTAFold) models; the change is chiefly the MSA search and packaging.
• “ColabFold runs AlphaFold3.” The repository does not say this. It lists an “AlphaFold3 (OpenFold3)” notebook and an --af3-json option that exports MSAs in an AlphaFold3-compatible JSON format.
• “A prediction is a measured structure.” It is a model prediction and should be treated as such.

FAQ

What is the difference between AlphaFold and ColabFold?

AlphaFold2 is the model presented by Demis Hassabis and John Jumper; ColabFold is an open-source system that runs AlphaFold2 (or RoseTTAFold) with an MMseqs2-based alignment search, where DeepMind’s notebook uses jackhmmer. The paper reports a 40-60-fold faster search.

Is ColabFold free to use?

The Nature Methods abstract describes ColabFold coupled with Google Colaboratory as a free and accessible platform, and the software is MIT licensed. Colab’s GPU availability and limits are set by Google and can change, so they are not guaranteed.

Is protein folding really solved?

The Royal Swedish Academy of Sciences’ press release for the 2024 Nobel Prize in Chemistry says Demis Hassabis and John Jumper presented AlphaFold2 in 2020 and that it has been used to predict the structure of virtually all the 200 million proteins researchers have identified. That is structure prediction from sequence. Whether a given prediction is right for your protein, and how proteins fold as a process, are separate questions that still need confidence metrics and experiments.

Is Google still working on AlphaFold?

Google DeepMind’s AlphaFold page still presents the work as active. It describes AlphaFold 3, introduced with Isomorphic Labs on 8 May 2024, and AlphaFold Server, launched the same day to give scientists free access to AlphaFold 3 predictions for non-commercial research. It also lists the AlphaFold Protein Structure Database, built with EMBL-EBI, as offering over 200 million predictions, and says AlphaFold is used by over 3 million researchers in over 190 countries (page text dated November 2025). Plans beyond what that page states are not something this guide can confirm.

Does ColabFold work with AlphaFold3?

The repository offers an AlphaFold3-compatible MSA export and an OpenFold3 notebook; it does not say ColabFold runs DeepMind’s AlphaFold3 model.

Sources

• Mirdita et al., “ColabFold: making protein folding accessible to all,” Nature Methods 2022;19:679-682 (PMID 35637307), Europe PMC record read 2026-11-08.
• sokrypton/ColabFold repository, github.com/sokrypton/ColabFold, read 2026-11-08.
• The Royal Swedish Academy of Sciences, press release, Nobel Prize in Chemistry 2024, nobelprize.org, read 2026-11-08.
• Google DeepMind, AlphaFold, deepmind.google/science/alphafold, read 2026-11-08.