ColabFold (https://github.com/sokrypton/ColabFold ) is an emerging protein folding prediction tool based on Google DeepMind’s Alphafold (see Using AlphaFold on O2 ). LocalColabFold (https://github.com/YoshitakaMo/localcolabfold )is a packaging of ColabFold for use on local machines; we provide instructions on how to leverage LocalColabFold on O2 below. LocalColabFold uses MMseqs2 (conditionally faster than jackhmmer), and runs AlphaFold2 for single protein modeling and AlphaFold-Multimer for protein complex modeling. If you are unsure about which to use, feel free to try both tools and compare results.
Note: If you’re new to Slurm or O2, please see Using Slurm Basic for lots of information on submitting jobs.
Using the Module
LocalColabFold is available via our LMOD module system (Using Applications on O2 ). As development of ColabFold is relatively fast, we have labeled the last updated time stamp in the help output of the module. We cannot guarantee that this module will be updated in a timely manner, so if you are interested in keeping up to date, please refer to the Installing LocalColabFold Locally section to learn how to set up a local installation you can maintain.
Presently, the module is hidden while we work out additional dependencies of the module. Though it presently still “works”, we do not currently recommend processing a large volume of proteins, as the default behavior of the tool is to send proteins to the remote server for alignment before returning to leverage the local GPU resource. When this nuance is resolved, this page will be updated accordingly; see the Caveats section for more information.
To access the module:
$ module load localcolabfold/latest
A snapshot of the help text follows:
$ module help localcolabfold/latest --------------------------- Module Specific Help for "localcolabfold/.latest" ----------------------------For detailed instructions, go to: https://github.com/YoshitakaMo/localcolabfold This module was last installed on February 9, 2022 using the latest colabfold commit as of ~4:30pm ET. Due to frequency in development updates, this module may be reinstalled any time. Please refer to the timestamp above for most recent installation time. This module currently requires gcc/9.2.0 to be loaded due to requiring external cuda libraries. If you are working under a different compiler stack (e.g. gcc/6.2.0), you may want to install this yourself until we offer an updated version of the cuda module under a different compiler. Visit the repository website for more information about how to install this yourself.
This output shows the last time the module was updated. If you are looking for a bleeding edge version of ColabFold, it may be preferable for you to install it yourself and manually keep your installation up to date with developments that are pertinent to your work.
Installing LocalColabFold Locally
If you would like to maintain control over when LocalColabFold is updated, you may choose to install it to a local folder under your direct control. We outline instructions on how to do so below.
First, begin an interactive session (https://harvardmed.atlassian.net/wiki/spaces/O2/pages/1586793632/Using+Slurm+Basic#Interactive-Sessions ) and load the conda module (preferably with nothing else loaded, run module purge to remove all active modules from your current environment):
$ module load miniconda3/4.10.3
Then, cd to a location where you would like to clone the LocalColabFold repository (https://github.com/YoshitakaMo/localcolabfold ), then do so:
$ git clone https://github.com/YoshitakaMo/localcolabfold.git
This command will create a folder in your working directory called localcolabfold. cd into this directory. Then, if you ls this location, you should see the files that are in the repository view on GitHub accessible via your terminal in your local folder that was just created. We are interested in the install_colabbatch_linux.sh file. Remember the path to this file (we will refer to this path as /path/to/install_colabbatch_linux.sh from now on; replace it with your own path). Now, cd to a location where you would like the environment to live, then invoke this script:
$ cd /path/to/desired/location $ sh /path/to/install_colabbatch_linux.sh
This should create a directory at /path/to/desired/location called colabfold_batch. Wait for the installation to finish.
Keeping Your LocalColabFold Installation Updated
The main reason you are probably maintaining a personal LocalColabFold installation is likely because you would like to use the newest features as they are implemented into ColabFold, without waiting for the module to be updated erratically (or perhaps not at all depending on stability). To keep your installation updated, return to the location where you cloned the repository (recall that it was called localcolabfold when it was created via git clone:
$ cd /path/to/localcolabfold
Now that you’re here, you’ll want to make sure that these scripts are updated:
$ git pull
Now, you’ll want to run the update_linux.sh script, and pass the path to your environment as an argument:
$ sh update_linux.sh /path/to/desired/location/colabfold_batch
After this is complete, you should have the newest version of ColabFold baked into your LocalColabFold installation.
Executing LocalColabFold Locally
Now that you have installed LocalColabFold locally, there is one more hurdle to using this installation.
First, make sure the bin subdirectory is added to your PATH variable:
$ export PATH=/path/to/desired/location/colabfold_batch/bin:$PATH
If you prefer, you can paste this line into your ~/.bashrc file instead so that it is automatically set up each time you log in to O2.
In order to leverage GPU resources, you will need to load a CUDA module, which in turn requires you to load a GCC module:
$ module load gcc/9.2.0 cuda/11.2
Via internal testing, Research Computing has discovered that even though install_colabbatch_linux.sh installs a local copy of (GCC and) CUDA, it is unable to leverage these resources, and we were unable to provide access to these local copies. The present workaround is to use the O2 modules in their place as above.
Executing LocalColabFold On O2
LocalColabFold can be loaded as a module by running:
module load gcc/9.2.0 cuda/11.2 localcolabfold/latest
or, if LocalColabFold has been installed locally, please make sure it is visible in your PATH variable for loading. Once you have loaded these modules, you’ll want to submit your job to the gpu (or gpu_quad if you have access) partition so that you can leverage GPU resources (Using O2 GPU resources ). Parameters for using LocalColabFold through the command colabfold_batch can be shown by loading the modules above in an interactive session (https://harvardmed.atlassian.net/wiki/spaces/O2/pages/1586793632/Using+Slurm+Basic#Interactive-Sessions ) and running:
colabfold_batch -h
At the moment, we do not recommend invoking ‘--amber’ or ‘--templates’ flags since these lead some jobs to fail
LocalColabFold accepts FASTA files containing amino acid sequences, including complexes where proteins are separated with a colon (:). These FASTA files can contain a single sequence, or a “batch” of several proteins as input. The path to this file should be included in a colabfold_batch command. Below is an simplified example of a colabfold_batch command (graciously provided by the Center for Computational Biomedicine):
colabfold_batch --num-recycle 5 \ --model-type AlphaFold2-multimer \ --rank ptmscore \ /PATH/TO/INPUT.fasta \ /PATH/TO/OUTPUT/DIRECTORY
These commands can be combined in an sbatch (https://harvardmed.atlassian.net/wiki/spaces/O2/pages/1586793632/Using+Slurm+Basic#The-sbatch-command ) script. The resources required to complete an sbatch job may vary by structure and complexity. It is generally best to start with a relatively conservative request for resources, then increase as needed based on information from past jobs. This information can be found using commands like O2sacct (Get information about current and past jobs ). Below is a simplified example of an sbatch script that runs the file INPUT.fasta against colabfold_batch on the gpu partition:
#!/bin/bash #SBATCH --partition=gpu # Partition to run in #SBATCH --gres=gpu:1 # GPU resources requested #SBATCH -c 1 # Requested cores #SBATCH --time=0-12:00 # Runtime in D-HH:MM format #SBATCH --mem=25GB # Requested Memory #SBATCH -o %j.out # File to which STDOUT will be written, including job ID (%j) #SBATCH -e %j.err # File to which STDERR will be written, including job ID (%j) #SBATCH --mail-type=ALL # ALL email notification type #SBATCH --mail-user=<email_address> # Email to which notifications will be sent module load gcc/6.2.0 cuda/11.2 localcolabfold/latest colabfold_batch --num-recycle 5 \ --model-type AlphaFold2-multimer \ --rank ptmscore \ /PATH/TO/INPUT.fasta \ /PATH/TO/OUTPUT/DIRECTORY
The output directory will contain several .pdb, .json, and .png files for the predicted complex structure. This includes pLDDT and PAE metrics that assess the accuracy of each prediction. The 'best' ranked structure will be called {sample_id}_unrelaxed_rank_1_model_{i}.pdb .
Caveats
LocalColabFold is a repackaging of ColabFold for local use. This means that LocalColabFold requires all the same local hardware resources and connections that ColabFold would require (but without the Google Colab soft dependency). This includes the allowing of shipping the protein sequence to a remote server maintained by the ColabFold developers for processing during the alignment step. This server is shared by all users of ColabFold, and is not an HPC environment to our knowledge. This means that SUBMITTING LARGE BATCHES OF PROTEINS IS NOT RECOMMENDED AT THIS TIME, regardless of whether you are using the O2 module or your own installation on O2. Do note that we are unable to quantify “large”, as this is to the discretion of the system administrators maintaining the remote server.
Large volumes of submissions to the remote server may cause the submitting compute node’s IP address to be rate-limited, or even blacklisted, which will impact all users of LocalColabFold on O2 that land on that compute node. Furthermore, because there are a limited number of compute nodes with GPU resources, if volume is high enough, all of O2’s GPU compute nodes can easily be blacklisted in a short amount of time.
Research Computing is working on a method to perform the alignment step locally, but it may result in longer processing times per protein. This page will be updated when this method is ready for consumption.
Troubleshooting/FAQ
Errors with using --amber or --templates
As noted above, occasionally jobs will fail if either of the above flags are enabled - this is a known issue and requires action from the ColabFold developers. For now, simply resubmit the job without these flags, or if these functions are required for your work, you can also try submitting your sequences against Alphafold (and adjust your resource requirements accordingly).
Please contact rchelp@hms.harvard.edu with any questions regarding the module or troubleshooting the installation process that this section does not address or addresses insufficiently. Depending on the question, we may need to refer you to the developers, but we will do our best to assist.