2025-01-29: Updated Uniref30 database to 2023_02

2024-12-03: Alphafold3 available to all users.

See our alphafold3 documentation.

2024-11-18: Alphafold3 installation in progress.

2024-05-22: The Alphafold 3 model was published recently (Abramson et al)

A public, non-commercial Alphafold 3 server is available at https://golgi.sandbox.google.com/about. But unfortunately the code and model are not publicly available at this time. From the paper:

Code availability: AlphaFold 3 will be available as a non-commercial usage only server at https://www.alphafoldserver.com, with restrictions on allowed ligands and covalent modifications. Pseudocode describing the algorithms is available in the Supplementary Information. Code is not provided.

2024-02-06: Added standalone relax and afpdb2cif scripts to alphafold/2.3.2 module

2024-01-20: Database updates and new alias

PDB was updated to the current version and a new alphafold command was added as an alias to the awkwardly named run_singularity for.

2023-05-23: alphafold 2.3.2 becomes the default version.

• No changes to model parameters or databases.
• --run_relax has been replaced with --models_to_relax with a default of best. That means that only the best model (ranked_0.pdb) will be relaxed.
• The jax python module is no loger required to read the pickled results files

2023-05-10: updated mgnify to 2023-02

2023-02-13: alphafold 2.3.1 becomes the default version.

2023-02-10: alphafold 2.3.1 available.

Some notable changes (See release page for full details):

• Added new AlphaFold-Multimer models with better accuracy on large protein complexes.
• Added early stopping to recycling.
• Used bf16 in multimer inference – reduces GPU memory usage.
• Relaxation metrics are now saved in relax_metrics.json.

2023-02-09: In place database update. This will apply to all alphafold version with the exception of the parameters which are version specific.

• uniref90 - 2022_05
• mgnify - 2022_05
• uniref30 - 2021_06 (unchanged since last update; this used to be called uniclast30)
• bfd - casp14 (unchanged since last update)
• uniprot - 2022_05
• pdb70 - 220313
• pdb mmcif - 230208
• pdb_seqres - 230208

2022-09-21: the --hhblits_extra_opts option was ported from msa to run_singularity

In a small number of cases hhblits fails to create alignments. This option can be used to fine tune the hhblits run (see below). Example: run_singularity --hhblits_extra_opts="-maxres 80000 -prepre_smax_thresh 50" ...

2022-07-11: the msa utility script has been disabled

Large scale use of the msa script may have been implicated in file system problems. The script has been removed until futher notice.

2022-06-02: added alphapickle to alphafold 2.2.0.

alphapickle will be included in alphafold installs ≥ 2.2.0

2022-04-22: Version 2.2.0 becomes the default

• See official announcement
• Switched to new AlphaFold-Multimer models with reduced numbers of clashes on average and slightly increased accuracy.
• removed --is_prokaryote_list option from run_singularity and msa as the prokaryotic pairing algorithm did not actually improve accuracy on average
• Added --num_multimer_predictions_per_model=N option to run_singularity. Runs N predictions per multimer model - each with different seeds. Defaults to 5 and will increase runtime
• Added --model_config option to run_singularity. This allows users to use a customized alphafold/model/config.py to alphafold for tuning certain parameters

2022-02-22: Version 2.1.2 becomes the default

• No changes to network weights
• Relaxation now defaults to using GPU
• Added options for controlling relaxation: --[no]run_relax and --[no]enable_gpu_relax
• New script to generate the multiple sequence alignments (MSAs) only (msa). This is highly recommended as MSA generation consumes about 60% of runtime and does not make use of GPU. The workflow is to run msa generation on CPU nodes and then run model prediction on a GPU node with run_singularity --use_precomputed_msas .... This script also supports adding extra options to hhblits or overriding some alphafold defaults with option --hhblits_extra_opts. Note that this script cannot run on x2650 nodes because it depends on an AVX2 hhblits build.

2021-11-15: Version 2.1.1 becomes the default

• the --preset option was renamed to --model_preset
• the --db_preset=reduced_dbs option is now supported for all alphafold versions
• The inofficial --use_ptm option became obsolete with introduction of the official --model_preset option and was removed
• This version uses the new model parameters (2021-10-27) released with 2.1.1. This includes the parameters for the multimer model. The 2.0.X modules will continue to use the previous parameters.

2021-11-14: Database update

Databases were updated in place: pdb mmcif and pdb70 (211110). New databases only used by multimer model: pdb_seqres, uniprot

2021-10-19: Added --use_ptm option to run_singularity

Use the pTM models, which were fine-tuned to produce pTM (predicted TM-score) and predicted aligned error values alongside their structure predictions.

2021-10-18: Adaptation of the alphafold_advanced notebook from ColabFold available in version 2.0.1.

Allows prediction of protein complexes with unmodified alphafold network weights. So far only an interactive notebook is available. See below for more details

2021-10-01: Version 2.0.0-24-g1d43aaf was tagged as 2.0.1

The modules for 2.0.0-24-g1d43aaf and 2.0.1 point to the same installation since the release was tagged after this revision was installed.

2021-09-21: Version 2.0.0-24-g1d43aaf becomes the default version on biowulf

Most noticable change should be the inclusion of pLDDT in the PDB B-factor column

2021-09-16: Database update (in place)

The following databases used by alphafold were updated in place: mgnify (2018_12 to 2019_05), pdb70 (200401 to 210901), pdb mmcif (210717 to 210915, 1969 additional structures), uniclust30 (2018_08 to 2021_06 from http://gwdu111.gwdg.de/~compbiol/uniclust/2021_06/). Uniref90 and BFD are unchanged.

• alphafold2 on GitHub
• Blog post

• Module Name: alphafold2 (see the modules page for more information)
• Alphafold2 first runs some multithreaded analyses using up to 8 CPUs before running model inference on the GPU. At this point these steps can't be separated and therefore for the first step of the job the GPU will remain idle.
• Example files in $ALPHAFOLD2_TEST_DATA
• Reference data in /fdb/alphafold2/
• Alphafold2 expects input to be upper case amino acid sequences

Allocate an interactive session and run the program. In this example the whole pipeline including multiple sequence alignment and model predictions are run with run_singularity on a GPU node.

[user@biowulf]$ sinteractive --mem=60g --cpus-per-task=8 --gres=lscratch:100,gpu:v100x:1
salloc.exe: Pending job allocation 46116226
salloc.exe: job 46116226 queued and waiting for resources
salloc.exe: job 46116226 has been allocated resources
salloc.exe: Granted job allocation 46116226
salloc.exe: Waiting for resource configuration
salloc.exe: Nodes cn3144 are ready for job

[user@cn3144]$ module load alphafold2/2.3.2

To predict the structure of a protein already in PDB without using its experimental structure as a template set max_template_date to before the release date of the structure. For example, to reproduce the T1049 CASP14 target with 144 aa. On a V100x this prediction runs for about 1h.

[user@cn3144]$ run_singularity --helpfull  # use --help for shorter help message
Singularity launch script for Alphafold.
flags:
/usr/local/apps/alphafold2/2.3.2/bin/run_singularity:
  --[no]benchmark: Run multiple JAX model evaluations to obtain a timing that excludes the compilation time,
    which should be more indicative of the time required for inferencing many proteins.
    (default: 'false')
  --db_preset: <full_dbs|reduced_dbs>: Choose preset MSA database configuration - smaller genetic
    database config (reduced_dbs) or full genetic database config (full_dbs)
    (default: 'full_dbs')
  --[no]dry_run: Print command that would have been executed and exit.
    (default: 'false')
  --[no]enable_gpu_relax: Run relax on GPU if GPU is enabled.
    (default: 'true')
  --fasta_paths: Paths to FASTA files, each containing a prediction target that will be folded one after
    another. If a FASTA file contains multiple sequences, then it will be folded as a multimer. Paths should
    be separated by commas. All FASTA paths must have a unique basename as the basename is used to name the
    output directories for each prediction. (a comma separated list)
  --gpu_devices: Comma separated list of devices to pass to NVIDIA_VISIBLE_DEVICES.
    (default: 'all')
  --max_template_date: Maximum template release date to consider (ISO-8601 format: YYYY-MM-DD). Important
    if folding historical test sets.
  --model_config: Use this file instead of default alphafold/model/config.py
  --model_preset: <monomer|monomer_casp14|monomer_ptm|multimer>: Choose preset model configuration -
    the monomer model, the monomer model with extra ensembling, monomer model with pTM head, or multimer model
    (default: 'monomer')
  --num_multimer_predictions_per_model: How many predictions (each with a different random seed) will be
    generated per model. E.g. if this is 2 and there are 5 models then there will be 10 predictions per
    input. Note: this FLAG only applies if model_preset=multimer
    (default: '5')
    (an integer)
  --output_dir: Path to a directory that will store the results.
  --[no]run_relax: Whether to run the final relaxation step on the predicted models. Turning relax off might
    result in predictions with distracting stereochemical violations but might help in case you are having
    issues with the relaxation stage.
    (default: 'true')
  --[no]use_gpu: Enable NVIDIA runtime to run with GPUs.
    (default: 'true')
  --[no]use_precomputed_msas: Whether to read MSAs that have been written to disk instead of running the
    MSA tools. The MSA files are looked up in the output directory, so it must stay the same between multiple
    runs that are to reuse the MSAs. WARNING: This will not check if the sequence, database or configuration
    have changed.
    (default: 'false')
...
absl.logging:
  --[no]alsologtostderr: also log to stderr?
    (default: 'false')
  --log_dir: directory to write logfiles into
    (default: '')
  --logger_levels: Specify log level of loggers. The format is a CSV list of `name:level`. Where `name` is the
    logger name used with `logging.getLogger()`, and `level` is a level name  (INFO, DEBUG, etc). e.g.
    `myapp.foo:INFO,other.logger:DEBUG`
    (default: '')
  --[no]logtostderr: Should only log to stderr?
    (default: 'false')
  --[no]showprefixforinfo: If False, do not prepend prefix to info messages when it's logged to stderr, --verbosity
    is set to INFO level, and python logging is used.
    (default: 'true')
  --stderrthreshold: log messages at this level, or more severe, to stderr in addition to the logfile.
    Possible values are 'debug', 'info', 'warning', 'error', and 'fatal'.  Obsoletes --alsologtostderr.
    Using --alsologtostderr cancels the effect of this flag. Please also note that this flag is
    subject to --verbosity and requires logfile not be stderr.
    (default: 'fatal')
  -v,--verbosity: Logging verbosity level. Messages logged at this level or lower will be included. Set to 1
    for debug logging. If the flag was not set or supplied, the value will be changed from the default of
    -1 (warning) to 0 (info) after flags are parsed.
    (default: '-1')
    (an integer)
...


[user@cn3144]$ run_singularity \
    --model_preset=monomer \
    --fasta_paths=$ALPHAFOLD2_TEST_DATA/T1049.fasta \
    --max_template_date=2022-12-31 \
    --output_dir=$PWD
###
### or use the equivalent alphafold alias
###
[user@cn3144]$ alphafold \
    --model_preset=monomer \
    --fasta_paths=$ALPHAFOLD2_TEST_DATA/T1049.fasta \
    --max_template_date=2022-12-31 \
    --output_dir=$PWD
[user@cn3144]$ tree T1049
T1049/
├── [user   1.1M]  features.pkl
├── [user   4.0K]  msas
│   ├── [user    33K]  bfd_uniclust_hits.a3m
│   ├── [user    18K]  mgnify_hits.sto
│   └── [user   121K]  uniref90_hits.sto
├── [user   170K]  ranked_0.pdb               # <-- shown below
├── [user   170K]  ranked_1.pdb
├── [user   170K]  ranked_2.pdb
├── [user   171K]  ranked_3.pdb
├── [user   170K]  ranked_4.pdb
├── [user    330]  ranking_debug.json
├── [user   170K]  relaxed_model_1.pdb
├── [user   170K]  relaxed_model_2.pdb
├── [user   170K]  relaxed_model_3.pdb
├── [user   170K]  relaxed_model_4.pdb
├── [user   171K]  relaxed_model_5.pdb
├── [user    11M]  result_model_1.pkl
├── [user    11M]  result_model_2.pkl
├── [user    11M]  result_model_3.pkl
├── [user    11M]  result_model_4.pkl
├── [user    11M]  result_model_5.pkl
├── [user    771]  timings.json
├── [user    87K]  unrelaxed_model_1.pdb
├── [user    87K]  unrelaxed_model_2.pdb
├── [user    87K]  unrelaxed_model_3.pdb
├── [user    87K]  unrelaxed_model_4.pdb
└── [user    87K]  unrelaxed_model_5.pdb

The processes prior to model inference on the GPU consumed up to 40 GB of memory for this protein. Memory requirements will vary with different size proteins.

The next example shows how to run a multimer model (available from version 2.1.1). The example used is a recently published PI3K structure.

[user@cn3144]$ cat $ALPHAFOLD2_TEST_DATA/pi3k.fa
>sp|P27986|P85A_HUMAN Phosphatidylinositol 3-kinase regulatory subunit alpha OS=Homo sapiens OX=9606 GN=PIK3R1 PE=1 SV=2
MSAEGYQYRALYDYKKEREEDIDLHLGDILTVNKGSLVALGFSDGQEARPEEIGWLNGYN
ETTGERGDFPGTYVEYIGRKKISPPTPKPRPPRPLPVAPGSSKTEADVEQQALTLPDLAE
QFAPPDIAPPLLIKLVEAIEKKGLECSTLYRTQSSSNLAELRQLLDCDTPSVDLEMIDVH
VLADAFKRYLLDLPNPVIPAAVYSEMISLAPEVQSSEEYIQLLKKLIRSPSIPHQYWLTL
QYLLKHFFKLSQTSSKNLLNARVLSEIFSPMLFRFSAASSDNTENLIKVIEILISTEWNE
RQPAPALPPKPPKPTTVANNGMNNNMSLQDAEWYWGDISREEVNEKLRDTADGTFLVRDA
STKMHGDYTLTLRKGGNNKLIKIFHRDGKYGFSDPLTFSSVVELINHYRNESLAQYNPKL
DVKLLYPVSKYQQDQVVKEDNIEAVGKKLHEYNTQFQEKSREYDRLYEEYTRTSQEIQMK
RTAIEAFNETIKIFEEQCQTQERYSKEYIEKFKREGNEKEIQRIMHNYDKLKSRISEIID
SRRRLEEDLKKQAAEYREIDKRMNSIKPDLIQLRKTRDQYLMWLTQKGVRQKKLNEWLGN
ENTEDQYSLVEDDEDLPHHDEKTWNVGSSNRNKAENLLRGKRDGTFLVRESSKQGCYACS
VVVDGEVKHCVINKTATGYGFAEPYNLYSSLKELVLHYQHTSLVQHNDSLNVTLAYPVYA
QQRR
>sp|P42336|PK3CA_HUMAN Phosphatidylinositol 4,5-bisphosphate 3-kinase catalytic subunit alpha isoform OS=Homo sapiens OX=9606 GN=PIK3CA PE=1 SV=2
MPPRPSSGELWGIHLMPPRILVECLLPNGMIVTLECLREATLITIKHELFKEARKYPLHQ
LLQDESSYIFVSVTQEAEREEFFDETRRLCDLRLFQPFLKVIEPVGNREEKILNREIGFA
IGMPVCEFDMVKDPEVQDFRRNILNVCKEAVDLRDLNSPHSRAMYVYPPNVESSPELPKH
IYNKLDKGQIIVVIWVIVSPNNDKQKYTLKINHDCVPEQVIAEAIRKKTRSMLLSSEQLK
LCVLEYQGKYILKVCGCDEYFLEKYPLSQYKYIRSCIMLGRMPNLMLMAKESLYSQLPMD
CFTMPSYSRRISTATPYMNGETSTKSLWVINSALRIKILCATYVNVNIRDIDKIYVRTGI
YHGGEPLCDNVNTQRVPCSNPRWNEWLNYDIYIPDLPRAARLCLSICSVKGRKGAKEEHC
PLAWGNINLFDYTDTLVSGKMALNLWPVPHGLEDLLNPIGVTGSNPNKETPCLELEFDWF
SSVVKFPDMSVIEEHANWSVSREAGFSYSHAGLSNRLARDNELRENDKEQLKAISTRDPL
SEITEQEKDFLWSHRHYCVTIPEILPKLLLSVKWNSRDEVAQMYCLVKDWPPIKPEQAME
LLDCNYPDPMVRGFAVRCLEKYLTDDKLSQYLIQLVQVLKYEQYLDNLLVRFLLKKALTN
QRIGHFFFWHLKSEMHNKTVSQRFGLLLESYCRACGMYLKHLNRQVEAMEKLINLTDILK
QEKKDETQKVQMKFLVEQMRRPDFMDALQGFLSPLNPAHQLGNLRLEECRIMSSAKRPLW
LNWENPDIMSELLFQNNEIIFKNGDDLRQDMLTLQIIRIMENIWQNQGLDLRMLPYGCLS
IGDCVGLIEVVRNSHTIMQIQCKGGLKGALQFNSHTLHQWLKDKNKGEIYDAAIDLFTRS
CAGYCVATFILGIGDRHNSNIMVKDDGQLFHIDFGHFLDHKKKKFGYKRERVPFVLTQDF
LIVISKGAQECTKTREFERFQEMCYKAYLAIRQHANLFINLFSMMLGSGMPELQSFDDIA
YIRKTLALDKTEQEALEYFMKQMNDAHHGGWTTKMDWIFHTIKQHALN

[user@cn3144]$ alphafold \
    --fasta_paths=$ALPHAFOLD2_TEST_DATA/pi3k.fa \
    --max_template_date=2021-11-01 \
    --model_preset multimer \
    --num_multimer_predictions_per_model=2 \
    --output_dir=$PWD
...snip...
[user@cn3144]$ exit

The model .pkl files which, unlike the .pdb files, are not re-ordered into ranked_ files, contain a lot of information about the models. These are python pickle files and python can be used to explore and visualize them. For example:

[user@cn3144]$ conda activate my_py39 # needs jupyter and the packages imported below
[user@cn3144]$ cd T1049
[user@cn3144]$ jupyter console
In [1]: import pickle
In [2]: import json
In [3]: import pprint
In [4]: import jax   # only needed for version 2.3.1
In [5]: pprint.pprint(json.load(open("ranking_debug.json", encoding="ascii")))
{'order': ['model_2_pred_0',
           'model_3_pred_0',
           'model_1_pred_0',
           'model_5_pred_0',
           'model_4_pred_0'],
 'plddts': {'model_1_pred_0': 88.44386138278787,
            'model_2_pred_0': 91.83564104655056,
            'model_3_pred_0': 88.49961929441032,
            'model_4_pred_0': 86.73066329994059,
            'model_5_pred_0': 87.4009420322368}}
### so model 2 is the best model in this run and corresponds to ranked_0.pdf
In [6]: best_model = pickle.load(open("result_model_2_pred_0.pkl", "rb"))
In [7]: list(best_model.keys())
Out[7]:
['distogram',
 'experimentally_resolved',
 'masked_msa',
 'predicted_lddt',
 'structure_module',
 'plddt',
 'ranking_confidence']
In [8]: best_model['plddt'].shape
Out[8]: (141,)

The predicted alignment error (PAE) is only produced by the monomer_ptm and multimer models. Since version 2.2.0 we also include alphapickle with alphafold to create plots, csv files, and chimera attribute files for each ranked model. By default output will be saved to the same folder. See -h for more options.

[user@cn3144]$ alphapickle -od T1049

If the model above was created with the monomer_ptm model the following two plots are generated for each model:

To get an idea of runtimes of alphafold2 we first ran 4 individual proteins on all our available GPUs. The proteins ranged in size from 144 aa to 622 aa. Note that for all but the smallest protein, K80 GPUs were not suitable and should not be considered for alphafold2. These tests were run with default settings except for a fixed --max_template_date=2021-07-31

The runtime to run all 4 protein on a V100x GPU with 8 CPUs and 60GB of memory was 3.2h, slightly less than the individual runtimes of the 4 proteins run separately. For this one job we also increased the number of CPUs to 16 or the number of GPUs to 2, neither of which appeared to shorted the runtime

The resource usage profile of the combined alphafold2 pipeline in our testing thus far is suboptimal and variable. Steps probably should be segregated into individual jobs with proper resources. We hope to optimize this in the future

#!/bin/bash
module load alphafold2/2.3.2
alphafold \
    --model_preset=monomer \
    --fasta_paths=$ALPHAFOLD2_TEST_DATA/T1049.fasta \
    --max_template_date=2020-05-14 \
    --output_dir=$PWD

Submit this job using the Slurm sbatch command.

sbatch --cpus-per-task=6 --partition=gpu --mem=60g --gres=gpu:v100x:1,lscratch:100 alphafold2_model.sh

Script for doing only the multiple sequence alignment.

#!/bin/bash
# this is msa.sh
module load alphafold2/2.3.2
alphafold \
    --model_preset=monomer \
    --fasta_paths=$ALPHAFOLD2_TEST_DATA/T1049.fasta \
    --max_template_date=2023-12-31 \
    --msas_only \
    --output_dir=$PWD

Script for doing only the model building using the sequence alignment created above.

#!/bin/bash
# this is model.sh
module load alphafold2/2.3.2
alphafold \
    --model_preset=monomer \
    --fasta_paths=$ALPHAFOLD2_TEST_DATA/T1049.fasta \
    --max_template_date=2023-12-31 \
    --use_precomputed_msas \
    --output_dir=$PWD

Submit these jobs using the Slurm sbatch command.

[user@biowulf]$ jobid=$(sbatch --cpus-per-task=6 --mem=60g --gres=lscratch:100 msa.sh)
[user@biowulf]$ echo $jobid
21865850
[user@biowulf]$ sbatch --cpus-per-task=6 --partition=gpu --mem=40g --gres=gpu:v100x:1,lscratch:100 --dependency=$jobid model.sh