Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 38 of the 40 references it cites were found in a registry or an index; 4 of the 19 sources its claims cite were checked for support.

ColabFold - Making protein folding accessible to all

Open access

Preprint — published as 10.1038/s41592-022-01488-1 · Paper page

Abstract

ColabFold offers accelerated protein structure and complex predictions by combining the fast homology search of MMseqs2 with AlphaFold2 or RoseTTAFold. ColabFold’s 40 - 60× faster search and optimized model use allows predicting close to a thousand structures per day on a server with one GPU. Coupled with Google Colaboratory, ColabFold becomes a free and accessible platform for protein folding. ColabFold is open-source software available at github.com/sokrypton/ColabFold . Its novel environmental databases are available at colabfold.mmseqs.com Contact: milot.mirdita@mpibpc.mpg.de , so@fas.harvard.edu , martin.steinegger@snu.ac.kr

Overview

CoChat Summary

TL;DR

ColabFold accelerates protein structure and complex prediction by replacing AlphaFold2/RoseTTAFold homology search with MMseqs2, making folding accessible via Google Colaboratory and local command-line tools. It reports 40–60-fold faster MSA generation, ~90-fold batch speed-up, and prediction accuracy matching AlphaFold2 on CASP14 and AlphaFold-multimer on ClusPro.

Key findings

  • ColabFold’s MMseqs2 search is 40–60-fold faster than AlphaFold2’s homology search, and batch prediction is ~90-fold faster by avoiding recompilation and adding early stopping; this enables close to 1,000 structures per day on one GPU.
  • On CASP14 free-modeling targets, mean TM-scores were 0.826 for ColabFold-AlphaFold2-BFD/MGnify, 0.818 for ColabFold-AlphaFold2-ColabFoldDB, 0.79 for AlphaFold2, 0.744 for AlphaFold-Colab, and 0.62 for ColabFold-RoseTTAFold-BFD/MGnify. Over all CASP14 targets, TM-scores were 0.887, 0.886, 0.888, and 0.754 for the respective methods excluding AlphaFold-Colab.
  • ColabFold achieved the highest accuracy for complex prediction on the ClusPro dataset with the AlphaFold-multimer model, though some targets performed better using residue-index mode.
  • The 1,762-protein *Methanocaldococcus jannaschii* proteome (proteins <1,000 amino acids) was predicted in 48 h on one Nvidia Titan RTX with early stopping at pLDDT ≥85; average pLDDTs for AlphaFold2 and ColabFold Stop ≥85 were 89.75 and 88.78 in a subsampled set of 50 proteins.

Methods

  • ColabFold combines an MMseqs2-based homology search server with AlphaFold2 or RoseTTAFold; it searches UniRef100/UniRef30, PDB70, and either BFD/MGnify or ColabFoldDB, using a diversity-aware filter and optional MSA pairing for complexes.
  • Benchmarks included 65 CASP14 targets containing 91 domains (20 free-modeling targets with 28 domains), 17 ClusPro complex targets evaluated with DockQ, and the *M. jannaschii* proteome; predictions were compared to experimental structures using TM-align and DockQ.
  • Run-time measurements separated MSA generation and model inference, using specified CPU/GPU systems; ColabFold modes were executed without templates for benchmarking.

Limitations

  • The paper states limitations of Google Colaboratory and provides the `colabfold_batch` command-line tool for tasks too large for Colab; it also notes AlphaFold-Colab cannot be used as a standalone and was excluded from the all-CASP14-target comparison.
  • Evident from the text: ColabFold benchmarks were run without templates, so template-enabled performance is not directly benchmarked; free-modeling CASP14 targets were used to optimize MMseqs2 search parameters, which may make those results optimistic.
  • Evident from the text: validation is limited to CASP14, ClusPro, and one proteome, so generalization to other target classes or larger proteomes is not established here.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1101/2021.08.15.456425

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Schematic diagram of ColabFold. a,b, ColabFold has a web and a command line interface (a) that send FASTA input sequence(s) to an MMseqs2 server (b) searching two databases, UniRef100 and a database of environmental sequences, with three profile-search iterations each. The second database is searched using a sequence profile generated from the UniRef100 search as input. The server generates two MSAs in A3M format containing all detected sequences. c, For predictions of single structures (i) we filter both A3Ms using a diversity-aware filter and return this to be provided as the MSA input feature to the AlphaFold2 models. For predictions of complexes (ii) we pair the top hits within the same species to resolve the inter-chain contacts and additionally add two unpaired MSAs (same as i) to guide the structure prediction. Single chain predictions are ranked by pLDDT and complexes by predicted TM-score. d, To help researchers judge the prediction quality we visualize MSA depth and diversity and show the AlphaFold2 confidence measures (pLDDT and PAE).
Fig. 1 Schematic diagram of ColabFold. a,b, ColabFold has a web and a command line interface (a) that send FASTA input sequence(s) to an MMseqs2 server (b) searching two databases, UniRef100 and a database of environmental sequences, with three profile-search iterations each. The second database is searched using a sequence profile generated from the UniRef100 search as input. The server generates two MSAs in A3M format containing all detected sequences. c, For predictions of single structures (i) we filter both A3Ms using a diversity-aware filter and return this to be provided as the MSA input feature to the AlphaFold2 models. For predictions of complexes (ii) we pair the top hits within the same species to resolve the inter-chain contacts and additionally add two unpaired MSAs (same as i) to guide the structure prediction. Single chain predictions are ranked by pLDDT and complexes by predicted TM-score. d, To help researchers judge the prediction quality we visualize MSA depth and diversity and show the AlphaFold2 confidence measures (pLDDT and PAE).CC BY · PubMed Central
Fig. 2 Comparison of predictions for single chains and complexes. a, Structure prediction comparison of AlphaFold2, AlphaFold-Colab and ColabFold-AlphaFold2 with BFD/MGnify and with the ColabFoldDB, and ColabFold-RoseTTAFold with BFD/MGnify using predictions of 91 domains of 65 CASP14 targets. The 28 domains from the 20 free-modeling (FM) targets are shown first. FM targets were used to optimize MMseqs2 search parameters. Each target was evaluated for each individual domain (in total 91 domains). b, MSA generation and model inference times for each CASP14 FM target sorted by protein length (same colors as before). Blue shows MSA run times for ColabFold-AlphaFold2-BFD/MGnify and ColabFold-RoseTTAFold-BFD/MGnify. c, Comparison of multimeric prediction modes in ColabFold and AlphaFold-multimer. The ColabFold modes include residue-index modification with models originally trained for single-chain predictions and those for multimeric prediction from AlphaFold-multimer, using DockQ (a quality measure for protein–protein docking models). d, Run time of colabfold_batch proteome prediction at three optimization levels: always recompile, default, and stop model/recycle evaluation after first prediction with a pLDDT of ≥85. The yellow dashed line represents an extrapolation on the basis of the 50 AlphaFold2 predictions. Source data
Fig. 2 Comparison of predictions for single chains and complexes. a, Structure prediction comparison of AlphaFold2, AlphaFold-Colab and ColabFold-AlphaFold2 with BFD/MGnify and with the ColabFoldDB, and ColabFold-RoseTTAFold with BFD/MGnify using predictions of 91 domains of 65 CASP14 targets. The 28 domains from the 20 free-modeling (FM) targets are shown first. FM targets were used to optimize MMseqs2 search parameters. Each target was evaluated for each individual domain (in total 91 domains). b, MSA generation and model inference times for each CASP14 FM target sorted by protein length (same colors as before). Blue shows MSA run times for ColabFold-AlphaFold2-BFD/MGnify and ColabFold-RoseTTAFold-BFD/MGnify. c, Comparison of multimeric prediction modes in ColabFold and AlphaFold-multimer. The ColabFold modes include residue-index modification with models originally trained for single-chain predictions and those for multimeric prediction from AlphaFold-multimer, using DockQ (a quality measure for protein–protein docking models). d, Run time of colabfold_batch proteome prediction at three optimization levels: always recompile, default, and stop model/recycle evaluation after first prediction with a pLDDT of ≥85. The yellow dashed line represents an extrapolation on the basis of the 50 AlphaFold2 predictions. Source dataCC BY · PubMed Central

Verification

Claims the paper makes

4 Valid · 4 Not valid · 12 Not checked · 51 Own results and interpretations

32 more claims

Create a free account to see all 71 claims and to verify the 15 references still open.

See all claims — free account
  • Probably not a substantial claim. 18
  • It may state two separate claims as one. 17
  • Borderline: one of the tests for a substantial claim is unclear. 1
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

7 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 40 references

Research works printed but not linked: 5

  1. Evans, R. et al. bioRxiv 2021.10.04.463034 (2021).
  2. (2020), Nucleic Acids Res., 48, D570
  3. Delmont, T. O. et al. bioRxiv 2020.10.15.341214 (2020).
  4. Alexander, H. et al. bioRxiv 2021.07.25.453713 (2021).
  5. Mosalaganti, S. et al. bioRxiv 2021.10.26.465776 (2021).

Cited by

See all 9,391

Metadata

Authors
Mirdita, M., Schütze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., & Steinegger, M.
Affiliations
  • Max Planck Institute for Multidisciplinary Sciences (DE)
  • Seoul National University (KR)
  • Collaborative Research Institute for Innovative Microbiology, The University of Tokyo (JP)
  • Michigan State University (US)
  • Harvard University (US)
Corresponding author
  • Mirdita, Milot
  • Ovchinnikov, Sergey
  • Steinegger, Martin
Venue
Nature Methods
Publisher
openRxiv
Published
2021-08-15
Volume
19
Issue
6
Pages
679-682
ISSN
1548-7091 (ISSN-L)
OpenAlex
W4281790889
Cited by
9,391
References
40
Citations by year
2026: 1465 · 2025: 2293 · 2024: 2715 · 2023: 2017 · 2022: 817 · 2021: 75
FWCI
644.88
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Nature Methods (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • Research Square (acceptedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • GoeScholar The Publication Server of the Georg-August-Universität Göttingen (Georg-August-Universität Göttingen) (submittedVersion, cc-by)
  • Seoul National University Open Repository (Seoul National University) (submittedVersion)
Funders
  • National Science Foundation (2032259)
  • Japan Agency for Medical Research and Development (JP21am0101107)
  • National Research Foundation (2019R1A6A1A10073437)
  • Bundesministerium für Bildung und Forschung
  • Seoul National University
  • Georg-August-Universität Göttingen
  • National Research Foundation of Korea (2021-R1C1-C102065, 2019R1A6A1A10073437, 2020M3A9G7103933, 2021R1C1C102065, 2021M3A9I4021220, 2021-M3A9-I4021220, 2020M3‐A9G7‐103933, 2019R1-A6A1-A10073437)
  • Max-Planck-Gesellschaft
  • National Institutes of Health (DP5OD026389)
  • National Institute of Allergy and Infectious Diseases (R21AI156595)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Structure and Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Structural Biology › Enzyme Structure and Function
Keywords
protein structure prediction, protein complex prediction, homology search, MMseqs2, AlphaFold2, RoseTTAFold, Google Colaboratory, GPU acceleration, open source software, environmental databases, ColabFold
MeSH terms
Computers, Proteins, Software, Databases, Factual, Protein Folding

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash and gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: ColabFold - Making protein folding accessible to all. https://app.cochat.ai/papers/10.1101/2021.08.15.456425
The paper
Mirdita, M., Schütze, K., Moriwaki, Y., Heo, L., Ovchinnikov, S., & Steinegger, M. (2021). ColabFold - Making protein folding accessible to all. Nature Methods. https://doi.org/10.1101/2021.08.15.456425

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.