ColabFold is available as a set of Jupyter notebooks that can be used on Google Colaboratory or on users’ local machines, as well as an easily installable command line application.
ColabFold has four main Jupyter notebooks24. The first is AlphaFold2_mmseqs2 for basic use, which supports protein structure prediction using MSAs generated by MMseqs2 (version edb822), custom MSA upload, use of template information, relaxing of the predicted structures using amber force fields25, and prediction of complexes. The second, AlphaFold2_advanced, for advanced users, additionally supports MSA generation using HMMer (same as AlphaFold-Colab), the sampling of diverse structures by iterating through a series of random seeds (num_samples), and control of AlphaFold2 model internal parameters, such as changing the number of recycles (max_recycle), the number of ensembles (num_ensemble), and the is_training option. The use of the is_training option enables dropout during inference. This activates the stochastic part of the model and can result in different predictions. Thus by iterating through different seeds, one can sample different structures predictions from the uncertainty of the model26 or the ambiguity of co-evolution constraints derived from the input MSA. The third main type of Jupyter notebook is AlphaFold2_batch, for batch prediction of multiple sequences or MSAs. The batch notebook saves time by avoiding recompilation of the AlphaFold2 models (section 2.5.2) for each individual input sequence. The fourth type is RoseTTAFold, for basic use of RoseTTAFold, and which supports protein structure prediction using MSAs generated by MMseqs2, and custom MSAs, and sidechain prediction using SCWRL4 (ref. 27). The RoseTTAFold notebook also has an option to use a slower but more accurate PyRosetta28 folding protocol for structure prediction, using constraints predicted by RoseTTAFold’s neural network.
ColabFold command line interface
We initially focused on making ColabFold as widely available as possible through our Notebooks running in Google Colaboratory. To meet the demand for a version that runs on local users’ machines, we released ‘LocalColabFold’. LocalColabFold can take command line arguments to specify an input FASTA file, an output directory, and various options to tweak structure predictions. LocalColabFold runs on a wide range of operating systems, such as Windows 10 or later (using Windows Subsystem for Linux 2), macOS and Linux. The structure inference and energy minimization are accelerated if a CUDA 11.1 or later compatible GPU is present. LocalColabFold is available as free open-source software at https://github.com/YoshitakaMo/localcolabfold
Recognizing the limitations of Google Colaboratory, we provide the colabfold_batch command line tool through the colabfold python package. This enables the computing of tasks on the user’s own computer that would have been too large for Google Colab, for example, predicting an entire proteome (Methods 2.7.4). It can be installed using Python’s pip package manager following the instructions at https://github.com/sokrypton/ColabFold. It can be used as colabfold_batch input_file_or_directory output_directory, supporting FASTA, A3M and CSV files as input.
Faster MSA generation with MMseqs2
Generating MSAs for AlphaFold2 and RoseTTAFold is a time-consuming task. To improve their run time while maintaining a high prediction accuracy, we implemented optimized workflows using MMseqs2.
MSA generation by MMseqs2
ColabFold sends the query sequence to an MMseqs2 server11. It searches the sequence(s) with three iterations against the consensus sequences of the UniRef30, a clustered version of the UniRef100 (ref. 29). We accept hits with an E-value lower than 0.1. For each hit we realign its respective UniRef100 cluster member using the profile generated by the last iterative search, filter them (Methods 2.2.2) and add these to the MSA. This expanding search results in a speed-up of ~10-fold given that only 29.3 million cluster consensus sequences are searched instead of all 277.5 million UniRef100 sequences. Additionally, it has the advantage of being more sensitive given that the cluster consensus sequences are used. We use the UniRef30 sequence profile to perform an iterative search against the BFD/MGnify or ColabFoldDB using the same parameters, filters and expansion strategy.
New diversity aware filter
To limit the number of hits in the final MSA we use the HHblits (v3.3.0) diversity filtering algorithm8 implemented in MMseqs2 in multiple stages. In the first stage, during UniRef cluster expansion, we filter each individual UniRef30 cluster before adding the cluster members to the MSA, such that no cluster pair has a higher maximum sequence identity than 95% (–max-seq-id 0.95). In the second stage, after realignment we enable only the –qsc 0.8 threshold and disable all other thresholds (–qid 0 –diff 0 –max-seq-id 1.0). Additionally, the qsc filtering is used only if at least 100 hits are found (–filter-min-enable 100). In the last stage, during MSA construction we filter again with the following parameters: –filter-min-enable 1000 –diff 3000 –qid 0.0,0.2,0.4,0.6,0.8,1.0 –qsc 0 –max-seq-id 0.95. Here, we extended the HHblits filtering algorithm to filter within a given sequence identity bucket such that it cannot eliminate redundancy across filter buckets. Our filter keeps the 3,000 most diverse sequences in the identity buckets [0.0–0.2], (0.2–0.4], (0.4–0.6], (0.6–0.8] and (0.8–1.0]. In buckets containing less than 1,000 hits we disable the filtering.
New MMseqs2 pre-computed index to support expanding cluster members
MMseqs2 was initially built to perform fast many-against-many sequence searches. Mirdita et al.11 improved it to also support fast single-against-many searches. This type of search requires the database to be indexed and stored in memory. mmseqs createindex indexes the sequences and stores all time-consuming-to-compute data structures used for MMseqs2 searches to disk. We load the index into the operating systems cache using vmtouch (https://github.com/hoytech/vmtouch) to enable calls to the different MMseqs2 modules to become nearly overhead free. We extended the index to store, in addition to the already present cluster consensus sequences, all member sequences and the pairwise alignments of the cluster representatives to the cluster members. With these resident in cache, we eliminate the overhead of the remaining module calls.
AlphaFold2 requires more than 2 TB of storage space for its databases, which is a significant hurdle for many researchers. We optimized its databases and additionally created another large environmental sequence database.
Reducing the size of BFD/MGnify
To keep all required sequences and data structures in memory we needed to reduce the size of the environmental databases BFD and MGnify, given that both databases together would have required ~517 GB RAM for headers and sequences alone.
BFD is a clustered protein database consisting of ~2.2 billion proteins organized in 64 million clusters. MGnify (2019_05) contains ~300 million environmental proteins. We merged both databases by searching the MGnify sequences against the BFD cluster-representative sequences using MMseqs2. Each MGnify sequence with a sequence identity of >30% and a local alignment that covers at least 90% of its length is assigned to the respective BFD cluster. All unassigned sequences are clustered at 30% sequence identity and 90% coverage (–min-seq-id 0.3 -c 0.3 –cov-mode 1 -s 3) and merged with the BFD clusters, resulting in 182 million clusters. To reduce the size of the database we filtered each cluster, keeping only the 10 most diverse sequences using mmseqs filterresult –diff 10. This reduced the total number of sequences from 2.5 billion to 513 million, thus requiring only 84 GB RAM for headers and sequences.
We built ColabFoldDB by expanding BFD/MGnify with metagenomic sequences from various environments. To update the database we searched the proteins from the SMAG (eukaryotes)14, MetaEuk (eukaryotes)13, TOPAZ (eukaryotes)15, MGV (DNA viruses)16, GPD (bacteriophages)17 and an updated version of MetaClust18 against the BFD/MGnify centroids using MMseqs2 and assigned each sequence to the respective cluster if they have a 30% sequence identity at a 90% sequence overlap (-c 0.9 –cov-mode 1 –min-seq-id 0.3). All remaining sequences were clustered using MMseqs2 cluster -c 0.9 –cov-mode 1 –min-seq-id 0.3 and appended to the database. We remove redundancy per cluster by keeping the most 10 diverse sequences using mmseqs filterresult –diff 10. The final database consists of 209,335,865 million representative sequences and 738,695,580 members (see the Data Availability section for the input files). We provide the MMseqs2 search workflow used in the server (Methods 2.2.1) as a standalone script (colabfold_search).
AlphaFold2 searches with HHsearch through a clustered version of the PDB (PDB70, ref. 8) to find the 20 top ranked templates. To save time, we use MMseqs2 (ref. 10) to search against the PDB70 cluster representatives as a prefiltering step to find candidate templates. This search is also done as part of the MMseqs2 API call on our server. Only the top 20 target templates according to E-value are then aligned by HHsearch. The accepted templates are given to AlphaFold2 as input features. This alignment step is done in the ColabFold client and therefore it requires the subset of the PDB70 containing the respective HMMs. The PDB70 subset and the PDB mmCIF files are fetched from our server. For benchmarking, no templates are given to ColabFold.
Modeling protein complexes with ColabFold
ColabFold offers protein complex folding through the specialized AlphaFold-multimer model and through manipulation of the residue index3. Here, we show the steps that we took for ColabFold to produce accurate protein complex predictions.
Modeling of protein–protein complexes
We implemented two protein complex prediction modes in ColabFold. One is based on AlphaFold-multimer4 and the other is based on the manipulation of residue index in the original AlphaFold2 model. Baek et al.3 show that RoseTTAFold is able to model complexes despite being trained only on single chains. This is done by providing a paired alignment and modifying the residue index. The residue index is used as an input to the models to compute positional embedding. In AlphaFold2 we find the same to be true, although surprisingly the paired alignment is often not needed (Fig. 2c). AlphaFold2 uses relative positional encoding with a cap at (left|i-jright|ge 32), meaning that any pair of residues separated by 32 or more are given the same relative positional encoding. By offsetting the residue index between two proteins to be > 32, AlphaFold2 treats them as separate polypeptide chains. ColabFold integrates this for modeling complexes.
For homo-oligomeric complexes (Supplementary Fig. 4a) the MSA is copied multiple times for each component. Interestingly, it was found that providing a separate MSA copy (padding by gap characters to extend to other copies) works significantly better than concatenating from left to right.
For hetero-oligomeric complexes (Supplementary Fig. 4b), a separate MSA is generated for each component. The MSA is paired according to the chosen pair_mode (section 2.4.2). Given that pLDDT is useful only for assessing local structure confidence, we use the fine-tuned model parameters to return the PAE for each prediction. As illustrated in Supplementary Fig. 4c, the inter-PAE, the predicted TM-score or interface TM-score (both derived from PAE) can be used to rank and assess the confidence of the predicted protein–protein interaction.
MSA pairing for complex prediction
A paired MSA helps AlphaFold2 to predict complexes more accurately only if orthologous genes are paired with each other. We followed a similar strategy as Bryant et al.22 to pair sequences according to their taxonomic identifier. For the pairing we search each distinct sequence of a complex against the UniRef100 using the same procedure as described in section 2.2.1. We return only hits that cover all complex proteins within one species and pair only the best hit (smallest E-value) with an alignment that covers the query to at least 50%. The pairing is implemented in the new MMseqs2 module pairaln.
For prokaryotic protein prediction we additionally implemented the protocol described in ref. 3 to pair sequences based on their distances in the genome as predicted from the UniProt accession numbers.
Taxonomic labels for MSA pairing
To pair MSAs for complex prediction, we retrieve for each found UniRef100 member sequence the taxonomic identifier from the NCBI (National Center for Biotechnology Information) Taxonomy database30. The taxonomic labels are extracted from the lowest common ancestor field (‘common taxon ID’) of each UniRef100 sequence from the uniref100.xml (2021_03) file.
Speeding up AlphaFold2’s model evaluation
Our efforts in speeding up AlphaFold2’s MSA generation yielded large improvements in its run time. However, we discovered multiple opportunities within AlphaFold2 to speed up its model inference without sacrificing (or only sacrificing very little of) its accuracy.
Avoid recompiling AlphaFold2 models
The AlphaFold2 models are compiled using JAX31 to optimize the model for specific MSA or template input sizes. When no templates are provided, we compile once and, during inference, replace the weights from the other models, using the configuration of model 5. This saves 7 min of compile time. When templates are enabled, model 1 is compiled and weights from model 2 are used, model 3 is compiled and weights from models 4 and 5 are used. This saves 5 min of compile time. If the user changes the sequence or settings without changing the length or number of sequences in the MSA, the compiled models are reused without triggering recompilation.
Avoid recompiling during batch computation
To avoid AlphaFold2 model recompilation for every protein AlphaFold2 provides a function to add padding to the input MSA and templates called make_fixed_size. However, this is not exposed in AlphaFold2. We used the function in our batch notebook as well as in our command line tool colabfold_batch, to maximize GPU use and minimize the need for model recompilation. We sort the input queries by sequence length and process them in ascending order. We pad the input features by 10% (by default). All sequences that lie within the query length and an additional 10% margin are not required to be recompiled, resulting in a large speed-up for short proteins.
AlphaFold2 improves the predicted protein structure by recycling (by default) three times, meaning that the prediction is fed multiple times through the model. We exposed the recycle count as a customizable parameter given that additional recycles can often improve a model (Supplementary Fig. 6) at the cost of a longer run time. We also implemented an option to specify a tolerance threshold to stop early. For some designed proteins without known homologous sequences, this helped to fold the final protein (Supplementary Fig. 5).
Speed-up of predictions through early stop
AlphaFold2 computes five models through multiple recycles. We noted that for prediction of high certainty (pLDDT >85), all five models would often produce structures of very similar confidence, for some even without or with less than three recycles. To speed up the computation we added a parameter to define an early stop criterion that halts additional model inferences and stops recycling if a given pLDDT or (interface) predicted TM-score threshold is reached.
In our investigation of AlphaFold2’s internal parameters we realized that we could expose many of the internal parameters that might be useful to researchers trying to explore AlphaFold2’s full potential.
Sampling of diverse structures
To reduce memory requirements, only a subset of the MSA is used as input to the model. Alphafold2, depending on model configuration, subsamples the MSA to a maximum of 512 cluster centers and 1,024 extra sequences. Changing the random seed can result in different cluster centers and thus different structure predictions. ColabFold provides an option to iterate through a series of random seeds, resulting in structure diversity. Further structure diversity can be generated by using the original or fine-tuned (use_ptm) model parameters and/or enabling is_training to activate the stochastic (dropout) part of the model. Enabling the latter can be used to sample an ensemble of models for the uncertain parts of the structure prediction.
ColabFold enables researchers to upload their own MSAs. Any kind of alignment tool can be used to generate the MSA. The uploaded MSA can be provided in aligned FASTA, A3M, STOCKHOLM or Clustal format. We convert the respective MSA format into A3M format using the reformat.pl script from the HH-suite8.
Lightweight 2D structure renderer
For visualization, we developed a matplotlib32 compatible module for displaying the 3D ribbon diagram of a protein structure or complex. The ribbon can be colored by residue index (N to C terminus) or by a predicted confidence metric (such as pLDDT). For complexes, each protein can be colored by chain ID. Instead of using a 3D renderer, we instead use a 2D line plotting based technique. The lines that make up the ribbon are plotted in the order in which they appear along the z-axis. Furthermore, we add shade to the lines according to the z-axis. This creates the illusion of a 3D rendered graphic. The advantage over a 3D renderer is that the images are very lightweight, can be used in animations and saved as vector graphics for lossless inclusion in documents. Given that the 2D renderer is not interactive, we additionally included a 3D visualization option using py3Dmol33 in the ColabFold notebooks.
We show with multiple datasets that ColabFold does not sacrifice accuracy for its much faster run times.
Benchmark with CASP14 targets
We compared AlphaFold-Colab and AlphaFold2 (commit b88f8da) against ColabFold using all CASP14 (ref. 2) targets. ColabFold-AlphaFold2 (commit 2b49880) used UniRef30 (2021_03) (ref. 34) and the BFD/MGnify or ColabFoldDB. ColabFold-RoseTTAFold (commit ae2b519) was executed with papermill (https://github.com/nteract/papermill) using the PyRosetta protocol28. ColabFold-RoseTTAFold-BFD/MGnify and ColabFold-AlphaFold2-BFD/MGnify used the same MSAs. AlphaFold-Colab used the UniRef90 (2021_03), MGnify (2019_05) and the small BFD. AlphaFold2 used the full_dbs preset and default databases downloaded with the download_all_data.sh script. The 65 targets contain 91 domains, among these are 20 free-modeling targets with 28 domains. We compared the predictions against the experimental structures using TMalign (downloaded on 24 February 2021) (ref. 35).
Measuring run times for CASP14 benchmark
To provide more accurate run times we split the MSA generation and model inference measurements. MSA generation was repeated five times and the MSA generation times were averaged.
ColabFold was executed using colabfold_batch. The MMseqs2 server that computes MSAs for ColabFold has 2 × 14 core Intel E5-2680v4 central processing units (CPUs) and 768 GB RAM. Each generated MSA was processed by a single CPU core. Run times were computed from server logs.
AlphaFold2 MSA generation run times were measured by running AlphaFold2 without models (providing an empty string to the –model_names parameter) on the same 2 × 14 core Intel E5-2680v4 CPUs and 768 GB RAM system. The AlphaFold2 databases were stored on a software-RAID5 as implemented in Linux (mdadm) composed of six Samsung 970 EVO Plus 1 TB NVMe drives. Run times for AlphaFold2 were taken from the features entry of the timings.json file. For a fair comparison, AlphaFold2 was modified to allow HMMer and HHblits to access one CPU core.
All ColabFold and AlphaFold2 model inference run-time measurements were done on systems with 2 × 16 core Intel Gold 6242 CPUs with 192 GB RAM and 4x Nvidia Quadro RTX5000 GPUs. Only one GPU was used in each run.
ColabFold-RoseTTAFold-BFD/MGnify and ColabFold-AlphaFold2-BFD/MGnify used the same MSAs, and run times are shown only once.
AlphaFold-Colab was executed in the browser using a Google Colab Pro account. The times for the homology search were taken from the notebook output cell ‘Search against genetic databases’. The JackHMMer search uses eight threads.
We compare predictions of 17 ClusPro4,12 targets to their native structures using DockQ (commit 3735c16) (ref. 36). We used colabfold_batch (commit 45ad0e9) with BFD/MGnify in residue index manipulation and AlphaFold-multimer mode to predict structures. We use MSA pairing as described in section 2.4.2 and also add unpaired sequences. Models are ranked by predicted interface TM-score as returned by AlphaFold-multimer. The DockQ AlphaFold-multimer reference numbers were provided by R. Evans.
We predict the proteome of M. jannaschii. Of the 1,787 proteins, we exclude the 25 proteins longer than 1,000 residues, leaving 1,762 proteins of 268 amino acids in average length. With the colabfold_search wrapper to MMseqs2 we search against the ColabFoldDB (section 2.3.2) in 113 min on a system with an AMD EPYC 7402P 24-core CPU (no hyperthreading) and 512 GB RAM. MMseqs2 had a maximum resident set size of 308 GB during the search. We then predict the structures on a single Nvidia Titan RTX with 24 GB RAM in 46 h using only MSAs (no templates). For each query we stop early if any recycle iteration reaches a pLDDT of at least 85. Early stopping results in a speed-up of 3.7-fold compared with the default and 4.8-fold compared with always recompiling. AlphaFold2 (reduced_dbs) was run with the reduced_dbs preset and no template information was used. We changed the AlphaFold2 source code to utilize all CPU cores during the homology search.
AlphaFold2 (reduced_dbs, v2.1.1), ColabFold (commit f5d0cec) default and ColabFold Stop ≥ 85 have an average pLDDT of 90.68, 90.22 and 89.33, respectively, for 50 randomly sampled proteins. These are the same proteins that were used to extrapolate the run time of AlphaFold2. Over all predictions, the pLDDTs for the M. jannaschii proteome downloaded from the AlphaFoldDB, ColabFold default and ColabFold Stop ≥ 85 are 89.75, 89.38 and 88.77, respectively.
Software used for analysis
Benchmark data analysis and visualization were done with R/4.1.1, ggplot/3.3.5, cowplot/1.1.1 and lubridate/1.7.10. ColabFold-generated plots were made using matplotlib/3.1.3. TM-score analysis was done with TMalign/2021/02/24 and DockQ/3735c16.
Further information on research design is available in the Nature Research Reporting Summary linked to this article.