AlphaFold at TACC

Last update: June 24, 2024

AlphaFold is a protein structure prediction tool developed by DeepMind (Google). It uses a novel machine learning approach to predict 3D protein structures from primary sequences alone. In July 2021, the developers made the source code available on Github and published a Nature paper (supplementary information) describing the method. In addition to the software, AlphaFold depends on ~2.9 TB of databases and model parameters. Researchers interested in making protein structure predictions with AlphaFold are encouraged to follow the guide below, and use the databases and model parameters that have been prepared.

Installations at TACC

Table 1. Installations at TACC

System	What's Available
Frontera	AlphaFold: v2.3.2 Data: /scratch2/projects/bio/alphafold/2.3.2/data Examples: /scratch2/projects/bio/alphafold/2.3.2/examples Module: /scratch2/projects/bio/alphafold/modulefiles
Lonestar6	AlphaFold: v2.3.2 Data: /scratch/tacc/apps/bio/alphafold/2.3.2/data Examples: /scratch/tacc/apps/bio/alphafold/2.3.2/examples Module: /scratch/tacc/apps/bio/alphafold/modulefiles
Stampede3	AlphaFold: v2.3.2 Data: /scratch/tacc/apps/bio/alphafold/2.3.2/data Examples: /scratch/tacc/apps/bio/alphafold/2.3.2/examples Module: /scratch/tacc/apps/bio/alphafold/modulefiles

Running AlphaFold

Structure Prediction from Single Sequence

To perform 3-D protein structure prediction with AlphaFold, first upload a fasta-formatted protein primary sequence to your $WORK or $SCRATCH (recommended) space. Sample fasta sequences are provided in the machine-specific "Examples" paths listed in the table above. A valid fasta sequence might look like:

>sample sequence consisting of 350 residues
MTANHLESPNCDWKNNRMAIVHMVNVTPLRMMEEPRAAVEAAFEGIMEPAVVGDMVEYWN
KMISTCCNYYQMGSSRSHLEEKAQMVDRFWFCPCIYYASGKWRNMFLNILHVWGHHHYPR
NDLKPCSYLSCKLPDLRIFFNHMQTCCHFVTLLFLTEWPTYMIYNSVDLCPMTIPRRNTC
RTMTEVSSWCEPAIPEWWQATVKGGWMSTHTKFCWYPVLDPHHEYAESKMDTYGQCKKGG
MVRCYKHKQQVWGNNHNESKAPCDDQPTYLCPPGEVYKGDHISKREAENMTNAWLGEDTH
NFMEIMHCTAKMASTHFGSTTIYWAWGGHVRPAATWRVYPMIQEGSHCQC

Next, prepare a batch job submission script for running AlphaFold. Two different templates for different levels of precision are provided within the "Examples" paths listed in Table 1. above:

• full_dbs.slurm: higher precision (default)
• reduced_dbs.slurm: higher speed

See the AlphaFold documentation for more information on the speed vs. quality tradeoff of each preset. The example templates each need to be customized before they can be used. Copy the desired template to your $WORK or $SCRATCH space along with the input fasta file. After necessary customizations, a batch script for running the full databases on Lonestar6 may contain:

#!/bin/bash
# full_dbs.slurm
# -----------------------------------------------------------------
#SBATCH -J af2_full                   # Job name
#SBATCH -o af2_full.%j.out            # Name of stdout output file
#SBATCH -e af2_full.%j.err            # Name of stderr error file
#SBATCH -p gpu-a100                   # Queue (partition) name
#SBATCH -N 1                          # Total # of nodes 
#SBATCH -n 1                          # Total # of mpi tasks 
#SBATCH -t 12:00:00                   # Run time (hh:mm:ss)
#SBATCH -A my-project                 # Project/Allocation name 
# -----------------------------------------------------------------

# Load modules
module unload xalt
module use /scratch/tacc/apps/bio/alphafold/modulefiles
module load alphafold/2.3.2-ctr

# Run AlphaFold2
run_alphafold.sh --flagfile=$AF2_HOME/examples/flags/full_dbs.ff \
                 --fasta_paths=$SCRATCH/input/sample.fasta \
                 --output_dir=$SCRATCH/output \
                 --model_preset=monomer \
                 --max_template_date=2020-05-14 \
                 --use_gpu_relax=True

In the batch script, make sure to specify the partition (queue) (#SBATCH -p), node / wallclock limits, and allocation name (#SBATCH -A) appropriate to the machine you are running on. Also, make sure the path shown in the module use line matches the machine-specific "Module" path listed in the Table 1. above.

The flagfile is a configuration file passed to AlphaFold containing parameters including the level of precision, the location of the databases for multiple sequence alignment, and more. Flag files for all presets can be found in the 'Examples' directory, and typically they should not be edited. The other three parameters passed to AlphaFold should be customized to your input path / filename, desired output path, and the selection of models. The parameters are summarized in the following table:

Table 2. AlphaFold Parameter Settings

Parameter	Setting
--fasta_paths	# full path including filename to your test data =$SCRATCH/input/sample.fasta
--output_dir	# full path to desired output dir (/scratch filesystem recommended) =$SCRATCH/output
--model_preset	# control which AlphaFold model to run, options are: =monomer | =monomer_casp14 | =monomer_ptm | =multimer
--max_template_date	# control which structures from PDB are used =2020-05-14
--use_gpu_relax	# whether to relax on GPUs (recommended if GPU available) =True | =False

Once the input fasta sequence and customized batch job script are prepared, submit to the queue with:

login1$ sbatch <job_script>

e.g.:

login1$ sbatch full_dbs.slurm

Using the scheme above with full_dbs precision, we expect each job to take between 2 to 12 hours depending on the length of the input fasta sequence, the speed of the compute node, and the relative load on the file system at the time of run. Using reduced_dbs should cut the job time in half, while slightly sacrificing precision. Refer to the AlphaFold Documentation for a description of the expected output files.

Batch Structure Predictions from Independent Sequences

To perform 3-D protein structure prediction with AlphaFold for many protein sequences, we recommend using TACC's Pylauncher utility. First review the instructions for submitting single sequence predictions above, then make the following adjustments:

Fasta formatted sequences should be uniquely identifiable either by giving each a unique name or by putting each sequence in its own uniquely-named directory. The simplest way to achieve this is to have one sub directory (e.g. $SCRATCH/inputs/) with all uniquely named fasta sequences in it:

login1$ ls $SCRATCH/inputs/
seq1.fasta
seq2.fasta
seq3.fasta
...

Next, prepare a launcher jobfile that contains each command that needs to be run. There should be one line in the jobfile for each input fasta sequence. Each line should refer to a unique input sequence and a unique output path:

Sample Alphafold Launcher Job File { #running-independentsequences-jobfile }

apptainer exec --nv $AF2_HOME/images/alphafold_2.3.2.sif /app/run_alphafold.sh --flagfile=$AF2_HOME/examples/flags/full_dbs.ff --fasta_paths=$SCRATCH/input/seq1.fasta --output_dir=$SCRATCH/output1 --model_preset=monomer --max_template_date=2020-05-14 --use_gpu_relax=True
apptainer exec --nv $AF2_HOME/images/alphafold_2.3.2.sif /app/run_alphafold.sh --flagfile=$AF2_HOME/examples/flags/full_dbs.ff --fasta_paths=$SCRATCH/input/seq2.fasta --output_dir=$SCRATCH/output2 --model_preset=monomer --max_template_date=2020-05-14 --use_gpu_relax=True
apptainer exec --nv $AF2_HOME/images/alphafold_2.3.2.sif /app/run_alphafold.sh --flagfile=$AF2_HOME/examples/flags/full_dbs.ff --fasta_paths=$SCRATCH/input/seq3.fasta --output_dir=$SCRATCH/output3 --model_preset=monomer --max_template_date=2020-05-14 --use_gpu_relax=True
`...`

NOTE: Due to the way launcher_gpu distributes tasks to individual GPUs, the full apptainer command must be used in thejobfile as shown above.

Prepare a batch job submission script by merging the AlphaFold template with a launcher template. Adjust the number of nodes, number of tasks, and the wall clock time appropriately for the number of jobs in the jobfile. For example, to run AlphaFold against six independent input sequences across two nodes (three per node) simultaneously, the job script would resemble:

#!/bin/bash
# full_dbs_launcher.slurm
# -----------------------------------------------------------------
#SBATCH -J af2_launcher_job             # Job name
#SBATCH -o af2_launcher_job.%j.out      # Name of stdout output file
#SBATCH -e af2_launcher_job.%j.err      # Name of stderr output file
#SBATCH -p gpu-a100                     # Queue (partition) name
#SBATCH -N 2                            # Total # of nodes
#SBATCH -n 6                            # Total # of mpi tasks
#SBATCH -t 16:00:00                     # Run time (hh:mm:ss)
#SBATCH -A my-project                   # Project/Allocation name
# -----------------------------------------------------------------

# Load modules (example path on Frontera)
module unload xalt
module use /scratch2/projects/bio/alphafold/modulefiles
module load alphafold/2.3.2-ctr

# Launcher specifics (use launcher_gpu for GPUs)
module load launcher_gpu
export LAUNCHER_JOB_FILE=jobfile

# Run AlphaFold with Launcher
${LAUNCHER_DIR}/paramrun

Once the input sequences, the jobfile, and the batch job submission script are all prepared, submit the job to the queue with:

login1$ sbatch <name_of_job_script>

e.g.:

login1$ sbatch full_dbs_launcher.slurm

Structure Prediction from Multiple Sequences (Multimer)

Nevertheless, we provide example flag files, job scripts, and sequences in the "Examples" paths listed above to test multimer folding. In our experience, the success rates of multimer folding jobs decrease as input sequence length increases.

References

• Github: AlphaFold
• AlphaFold Nature Paper