henhaobio
Sign in

Nature MethodsBioinformaticsPeer-reviewed journal article

ColabFold - Making protein folding accessible to all

Milot MirditaiD, Konstantin SchützeiD, Yoshitaka MoriwakiiD, Lim HeoiD, Sergey OvchinnikoviD, Martin SteineggeriD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

ColabFold offers accelerated protein structure and complex predictions by combining the fast homology search of MMseqs2 with AlphaFold2 or RoseTTAFold. ColabFolds 40 - 60x 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 Contactmilot.mirdita@mpibpc.mpg.de, so@fas.harvard.edu, martin.steinegger@snu.ac.kr

In-depth summary

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text (15 Aug 2021) · CC BY 4.0

ColabFold couples 40–60-fold faster MMseqs2 homology search to AlphaFold2, matching it on CASP14 and AlphaFold-multimer on ClusPro complexes, yielding close to 1,000 structures daily per GPU.

Running AlphaFold2 or RoseTTAFold requires HMMer/HHblits multiple sequence alignment (MSA) searches that can take hours per protein over databases exceeding 2 TB, plus large-memory GPUs, and the free AlphaFold-Colab notebook loses accuracy with an eightfold-reduced database yet remains slow. ColabFold instead offers a hosted MMseqs2 server, compact environmental databases and free notebook or command-line prediction of monomers, complexes and proteomes.

Key results

  1. ShownTemplate-free ColabFold-AlphaFold2 matched full AlphaFold2 run with templates on CASP14 and beat the AlphaFold-Colab notebook on free-modeling targets; more recycles further helped targets with sparse MSAs.

    Mean template modeling (TM)-score over 65 CASP14 targets (91 domains): 0.887 ColabFold-AlphaFold2-BFD/MGnify, 0.886 with ColabFoldDB, 0.888 AlphaFold2, 0.754 ColabFold-RoseTTAFold; 20 free-modeling targets: 0.826 and 0.818 vs 0.79 AlphaFold2 and 0.744 AlphaFold-Colab. 12 recycles (default 3) raised the average to 0.898. · Main; Recycle count; Benchmark with CASP14 targets

  2. ShownReplacing HMMer/HHblits with MMseqs2 profile searches over clustered UniRef30, plus reusing compiled models, cuts end-to-end single-structure prediction time well below AlphaFold2 and AlphaFold-Colab.

    MMseqs2 search 40–60-fold faster than AlphaFold2's; single predictions average fivefold faster than AlphaFold2 and AlphaFold-Colab counting MSA generation plus inference; searching 29.3 million UniRef30 consensus sequences rather than 277.5 million UniRef100 sequences gives ~10-fold speed-up; compiling one model and swapping weights saves 7 min (5 min with templates). · Main; MSA generation by MMseqs2; Avoid recompiling AlphaFold2 models; Measuring run times for CASP14 benchmark

  3. ShownBatch mode (length-sorted padding to avoid recompilation, reused weights, early stopping on predicted local distance difference test (pLDDT) confidence) folded the Methanocaldococcus jannaschii proteome on a single GPU with little confidence loss.

    1,762 proteins under 1,000 residues (mean 268 aa): ColabFoldDB search 113 min, folding 46 h on one Nvidia Titan RTX; early stop at pLDDT ≥85 gave 3.7-fold speed-up vs default and 4.8-fold vs always recompiling; ~90-fold overall vs AlphaFold2 (extrapolated from 50 runs). Average pLDDT, all proteins: 89.75 AlphaFoldDB, 89.38 ColabFold default, 88.77 early stop. · Main; Avoid recompiling during batch computation; Speed-up of predictions through early stop; Proteome benchmark

  4. SuggestedColabFold models homo- and heteromeric complexes either with AlphaFold-multimer or by offsetting residue indices in single-chain AlphaFold2, using species-paired plus unpaired MSAs, and matched AlphaFold-multimer on a small docking benchmark.

    DockQ (docking model quality) vs native structures on 17 ClusPro targets: ColabFold's AlphaFold-multimer mode was most accurate, matching AlphaFold-multimer reference values, though some targets scored better with residue-index offsets (>32, past AlphaFold2's relative-position cap); models ranked by predicted interface TM-score. Residue-index mode often needed no pairing. · Main; Modeling of protein–protein complexes; MSA pairing for complex prediction; Complex benchmark

  5. ShownA redundancy-reduced merge of BFD and MGnify fits the environmental search set in memory, and ColabFoldDB adds eukaryotic, DNA-virus and phage protein catalogs that diversify MSAs for sparse families, though CASP14 accuracy did not improve.

    Merging BFD with MGnify 2019_05 into 182 million clusters and keeping the 10 most diverse sequences per cluster cut 2.5 billion to 513 million sequences and header/sequence RAM from ~517 GB to 84 GB. ColabFoldDB (209,335,865 representatives, 738,695,580 members) diversified MSAs for Pfam domains with <30 members; overall CASP14 TM-score 0.886 vs 0.887 with BFD/MGnify. · Main; Reducing the size of BFD/MGnify; ColabFoldDB

Methods, data and reproducibility

A server runs MMseqs2 profile searches (UniRef30 consensus expanded to UniRef100 members, then BFD/MGnify or ColabFoldDB), diversity-filters the MSAs and, for complexes, pairs them by taxonomy; these feed AlphaFold2, AlphaFold-multimer or RoseTTAFold in Colab notebooks or colabfold_batch, with compiled-model reuse and confidence-based early stopping.

  • All CASP14 targets scored per domain by TM-score against experimental structures; the 20 free-modeling targets (28 domains) also served to tune search parameters · n = 65 targets (91 domains)
  • ClusPro protein complex targets compared with native structures using DockQ · n = 17 targets
  • Methanocaldococcus jannaschii proteome (proteins <1,000 residues) for batch prediction; pLDDT compared with AlphaFold2 on 50 random proteins and with AlphaFoldDB models · Methanocaldococcus jannaschii · n = 1,762 of 1,787 proteins · UniProt UP000000805; AlphaFoldDB UP000000805_243232_METJA_v1
  • Pfam 34.0 seed and full alignments; families with <30 members used to compare MSA diversity from ColabFoldDB vs BFD/MGnify
  • UniRef30 (2021_03) cluster consensus sequences expanded to UniRef100 members, with NCBI Taxonomy IDs from uniref100.xml (2021_03) for MSA pairing · multiple species · n = 29.3 million consensus; 277.5 million UniRef100 sequences
  • BFD and MGnify (2019_05) merged via MMseqs2 (30% identity, 90% coverage), keeping the 10 most diverse sequences per cluster (BFD/MGnify) · multiple species (environmental sequences) · n = 513 million sequences; 182 million clusters
  • ColabFoldDB: BFD/MGnify plus eukaryotic (SMAG, MetaEuk, TOPAZ), DNA-virus (MGV), phage (GPD) and updated MetaClust proteins, clustered at 30% identity, 90% coverage · multiple species (eukaryotes, DNA viruses, bacteriophages) · n = 209,335,865 representatives; 738,695,580 members
  • PDB70 profile HMM database for templates (MMseqs2 prefilter, top 20 by E-value aligned with HHsearch); templates disabled in ColabFold benchmarks
Methods & tools
MMseqs2 (version edb822), incl. new pairaln module and colabfold_search workflow, AlphaFold2 (commit b88f8da; v2.1.1 reduced_dbs for proteome comparison), AlphaFold-multimer, RoseTTAFold (ColabFold commit ae2b519) with PyRosetta folding protocol and SCWRL4, AlphaFold-Colab notebook (HMMer/JackHMMer search), HH-suite (HHblits v3.3.0 filter algorithm, HHsearch, reformat.pl), JAX model compilation with make_fixed_size padding, TMalign (2021/02/24), DockQ (commit 3735c16), Amber force-field relaxation, py3Dmol and matplotlib-based 2D ribbon renderer, Google Colaboratory Jupyter notebooks and LocalColabFold
Software & environment
MMseqs2 edb822; AlphaFold2 commit b88f8da (v2.1.1 for proteome run); ColabFold commits 2b49880 (CASP14), 45ad0e9 (complexes), f5d0cec (proteome); LocalColabFold uses CUDA 11.1 or later GPUs if present; R/4.1.1, matplotlib/3.1.3.
Assay conditions & compute
MSA server: 2 × 14-core Intel E5-2680v4 CPUs, 768 GB RAM; inference on one Nvidia Quadro RTX5000 per run; proteome search 113 min on AMD EPYC 7402P (512 GB RAM, 308 GB peak), folding 46 h on one Nvidia Titan RTX (24 GB).
Reproducibility gaps
AlphaFold-multimer DockQ reference numbers were supplied by R. Evans rather than computed by the authors; notebooks rely on the authors' hosted MMseqs2 server, and local colabfold_search over ColabFoldDB peaked at 308 GB RAM.
Limitations (6)
  • Default MMseqs2 compositional filtering can hurt specific targets: CASP14 T1084 rose from TM-score 0.457 to 0.872 only when the filter was disabled, and other targets diverging notably from AlphaFold2 are listed in a supplementary table. · stated by the authors
  • The 20 free-modeling CASP14 targets (28 domains) were also used to tune MMseqs2 search parameters, which the authors flag by plotting them separately from the remaining targets. · stated by the authors
  • Google Colaboratory assigns machines from a pool of uneven capability and cannot handle jobs as large as a whole proteome, so large runs require local installation via colabfold_batch or LocalColabFold. · stated by the authors
  • No single complex mode wins everywhere: AlphaFold-multimer mode was best overall on ClusPro, yet some targets were better with residue-index offsets, and paired MSAs only help when true orthologs are matched. · stated by the authors
  • Complex-prediction parity with AlphaFold-multimer rests on 17 ClusPro targets, summarized graphically, with no per-target DockQ values, uncertainty estimates or significance tests in the text. · noted in this summary
  • Proteome runs exclude the 25 proteins over 1,000 residues and judge early stopping only by pLDDT, not experimental structures; AlphaFold2 vs early-stop pLDDT on 50 proteins differs between main text (89.75 vs 88.78) and Methods (90.68 vs 89.33). · noted in this summary

Read this if you need AlphaFold2 or RoseTTAFold predictions of monomers, complexes or whole proteomes without 2 TB databases or large servers, or want to tune MSA filtering, recycling and early stopping.

Citations
9,394
FWCI
644.14
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISRoberts '23Elliott '24Xu '19Li '22Lamb '21Lee '22
73% semantic matchBioinformatics
Jacob B. Roberts, Alberto A. Nava, Allison N. Pearson+5 more·bioRxiv·May 11, 2023· 4 citations