Protein Structure · Steinegger Lab
ColabFold
MMseqs2-accelerated AlphaFold2 protein monomer and complex structure prediction pipeline.
What ColabFold is
ColabFold is a protein structure and complex prediction system developed by the Steinegger Lab (created by Sergey Ovchinnikov, Milot Mirdita, and Martin Steinegger). It accelerates the prediction pipeline by replacing the standard, computationally intensive Jackhmmer homology search in AlphaFold2 with MMseqs2. This modification reduces multiple sequence alignment generation time from hours to seconds while maintaining structural accuracy.
The tool was designed to make structure prediction accessible and computationally practical for large batches of sequences. Alongside Google Colab notebooks, it provides command-line interfaces for both local and server-based multiple sequence alignment searches, batch predictions, and integration with downstream structural analysis tools.
ColabFold supports monomer and multimeric complex modelling using AlphaFold2 and AlphaFold-multimer, and includes features for amber relaxation, export to AlphaFold3-compatible formats, and accelerated execution via fused GPU kernels.
How it works
ColabFold pairs fast sequence search algorithms with deep learning-based structural folding models. For a given input sequence or set of sequences, it uses MMseqs2 to rapidly query large sequence databases (such as ColabfoldDB or UniRef30) to generate a multiple sequence alignment. Alternatively, users can send queries to a public MSA server or use an MMseqs2 GPU server for accelerated retrieval.
The generated alignment is fed into AlphaFold2 or AlphaFold-multimer network weights implemented in JAX. Complex prediction is supported either by introducing residue index jumps in standard AlphaFold2 or through AlphaFold-multimer models. The model outputs 3D coordinates in PDB format alongside per-residue predicted Local Distance Difference Test (pLDDT) confidence scores, which are stored in the B-factor column of the PDB output.
Key facts
- Tasks
- Structure prediction, complex prediction
- Input modalities
- Protein sequence (FASTA, CSV)
- Output modalities
- PDB structure coordinates, per-residue pLDDT vector
- Hardware requirement
- 40 GB GPU memory (up to ~2000 residues on ~16 GB GPUs)
- Alignment backend
- MMseqs2 (release 18)
- Supported architectures
- AlphaFold2, AlphaFold-multimer
What you can use it for
- Protein Monomer Structure Prediction
- Predict all-atom coordinates and per-residue pLDDT confidence scores from individual amino acid sequences using MMseqs2 alignments.
- Protein Complex Modelling
- Predict multimeric assemblies using either AlphaFold-multimer or AlphaFold2 with residue index jumps.
- Batch Alignment Preparation for AlphaFold3
- Generate multiple sequence alignments and export them into AlphaFold3-compatible JSON files, including specifications for non-protein molecules like DNA, RNA, ligands, and SMILES strings.
- Crystallographic Molecular Replacement Phasing
- Generate initial search models for crystallographic phasing, taking into account that the B-factor field contains pLDDT confidence values rather than true crystallographic B-factors.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Not runnable yet
This model isn't available to run right now.Run this model
ColabFold
Strengths and limits
Good at
- Cuts sequence alignment time from hours down to seconds compared to Jackhmmer without sacrificing prediction accuracy.
- Supports fast predictions using accelerated fused kernels (via Pallas or CUDA legacy backends) that yield roughly 2.5x faster runtime with reduced memory usage.
- Enables reuse of compilation costs across runs using persistent JAX compilation caching, padding, and query sorting by length or MSA depth.
- Provides GPU-accelerated MSA searches via MMseqs2-GPU and a dedicated GPU server for minimal latency.
Watch out for
- Predicting sequences on ~16 GB GPUs is limited to lengths of approximately 2000 residues.
- Batch searches with local databases require substantial system resources, needing around 940 GB of disk space and between 128 GB and over 1 TB of RAM depending on index preloading.
- MMseqs2 MSA generation in ColabFold only applies to protein sequences, leaving RNA alignments null in AlphaFold3-formatted exports.
- Output PDB files store pLDDT scores in the B-factor column, which requires recalibration or conversion before use in programs like Phenix Phaser that expect standard B-factors.
Inputs and outputs
You provide
Protein sequence
Use a colon to separate chains for complex prediction.
You get back
All-atom coordinates in PDB format, viewable and downloadable.
pLDDT score per residue (0–100). Above 70 is generally reliable.
Licence
Commercial use permitted · MIT
The MIT licence allows use in commercial research and products.
Questions
Can I use ColabFold commercially?
Commercial use permitted. The MIT licence allows use in commercial research and products.
How do I run ColabFold without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 1.5min, and a run costs about 280 credits. Visitors without an account get 1,000 credits a month.
What confidence metrics does ColabFold provide for its predicted structures?
It provides a per-residue predicted Local Distance Difference Test (pLDDT) score ranging from 0 to 100. Scores above 70 are generally considered reliable.
How are protein complexes predicted?
ColabFold supports complex prediction either through the dedicated AlphaFold-multimer model or through standard AlphaFold2 using a residue index jump.
Can ColabFold output files suitable for AlphaFold3?
Yes, using the --af3-json option, ColabFold can format protein MSAs and complex metadata—including nucleic acids, CCD codes, and SMILES strings—into an AlphaFold3-compatible JSON file.
Why does the first prediction take several minutes before folding begins?
JAX compiles an optimised model the first time it encounters a specific sequence length or alignment depth. This compile time can be mitigated on repeated runs by enabling the persistent JAX compilation cache.
How to cite it
Mirdita M, Schütze K, Moriwaki Y, Heo L, Ovchinnikov S and Steinegger M. ColabFold: Making protein folding accessible to all. Nature Methods (2022) doi: 10.1038/s41592-022-01488-1
Written from ColabFold's own documentation (code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.