AlphaFold2 standalone preinstalled on HPC@ISU

The Pronto and Nova clusters currently have three versions of AlphaFold pre-installed, which you can find using the module command:

module avail alphafold

---------------------------------------- /opt/rit/singularity/modules ----------------------------------------
   alphafold/2.0.0    alphafold/2.1.1    alphafold/2.1.2 (D)

  Where:
   L:  Module is loaded
   D:  Default Module

The latest version will be loaded by default (D), but you can select any of the available versions.

module load alphafold                   # loads alphafold/2.1.2
module load alphafold/2.1.1             # loads alphafold/2.1.1

To facilitate the preparation of AlphaFold jobs and submission into the SLURM queue, the HPC team has prepared a set of instructions and tips available via the execution of note_4_demo alias. The notes along with our benchmark results are included in the Performance of the AlphaFold section.

note_4_demo                                                            # prints contents on the screen
less note_4_demo                                                       # opens the preview of the
cp -r  /opt/rit/singularity/images/alphafold/demo ~/your_workdir/      # copies the demo/ with sample SLURM scripts

After loading the selected AlphaFold module, you can use the predefined alphafold wraps, which contain aliases to the corresponding option sets. More detailed options for each of the wraps can be viewed with the --help or --helpfull options.

  • alphafold , only single-chain modeling allowed
  • alphafold_casp14 , uses the settings from CASP14
  • alphafold_multimer , multi-chain structure modeling allowed

AlphaFold options

To display the full list of alphafold options, execute:

alphafold --help

The basic options you need to set up in each job submitted to the queue are:

option value description
–fasta_paths= /absolute/path/to/your/FASTA/input_dir/filename.fasta paths to input FASTA files
–output_dir= /absolute/path/to/your/outputs path to a directory that will store the results

Other options that may be useful:

flag description
–use_precomputed_msas If you job failed in the step after the sequence alignment, you can use this option to re-use previous sequence alignments and speed-up prediction.
–use_gpu_relax The relaxation stage will run on GPU. The flag is in all alphafold* wraps already. Not need to add in user’s script.
–norun_relax The relaxation stage can now be disabled using the run_relax flag. Then, you can quickly get a few unrelaxed structure models.

AlphaFold Inputs

Likewise with ColabFold, the standalone implementation of AlphaFold on HPC clusters also requires the amino acid sequence of the protein in FASTA format as the only input file.

  • In particular, alphafold wrap is intended to model monomeric proteins, that is, proteins with a single chain. Therefore, there can only be one protein sequence in the input file and you will need a separate FASTA files for all your targets.

    Input file: 2gb1.fasta or monomer.fa

 >2GB1_1|Chain A|PROTEIN G|Streptococcus sp.
 MTYKLILNGKTLKGETTTEAVDAATAEKVFKQYANDNGVDGEWTYDDATKTFTVTE

  • While, alphafold_multimer is designed to model multimeric proteins, that is, proteins with multiple chains. Therefore, there can only be one protein complex per sequence file, which has the multiple sequence entries > for each of the required chains.

    Input file: complex.fasta or multimer.fa

  >Chain_A
  MDEDGLPLMGSGIDLTKVPAIQQKRTVAFLNQFVVHTVQFLNRFSTVCEEKLADLSLRIQQ
  >CHAIN_B
  MNRTTPDQELVPASEPVWERPWSVEEIRRSSQSWSLAADAGLLQFLQEFSQQTISRTHEIKKQVDGLIRETKATDCRLHNVFNDFLMLSNTQFIENRVYDEEVEEPVL

###SLURM script configuration

By the principal rule, all calculations performed in the HPC infrastructure should be delegated to the computing nodes. To do this, you need to prepare a configuration script and submit the job to a queue. The task manager on the HPC@ISU infrastructure is SLURM.

Parameters required to set up AlphaFold tasks

Protein structure modeling with AlphaFold is a multi-operation process. For example, for sequence alignment and template search AlphaFold uses CPUs but other computational steps, such as amber-relaxation, can be accelerated by using available GPUs. Newer versions of AlphaFold can recognize more than one GPU card when starting but only one card is used in the tensorflow-GPU step. So, please allocate only 1 GPU card to avoid blocking resources for other users.

  • See AlphaFold performance section for details regarding content of the Table below

    param SMALL MEDIUM LARGE HUGE notes
    SEQ length <= 200aa <= 800aa <= 2500aa > 2500aa  
    #SBATCH          
    –nodes= 1 1 1 1 AlphaFold runs on single node. No MPI.
    –ntasks= 8 8* 8* 8* AlphaFold uses hard-coded 8 or 4 CPUs
    for sequence alignment and template search,
    so ntasks has to be 8 or larger.
    –mem= 180G 180G 180G 256G or 312G If you got run-out-of memory error,
    simply increase the memory allocation.
    –time= 1:0:0 5:0:0 10:0:0 1-0:0:0 Request extension if needed.
    -p gpu gpu gpu amdgpu  
    NOVA          
    –gres= gpu:1 gpu:v100:1 gpu:a100:1 gpu:a100:1 Allocate ONE GPU card.
    PRONTO          
    –gres= gpu:1 gpu:rtx_6000:1 gpu:a100-pcie:1 gpu:a100-pcie:1 Allocate ONE GPU card.

    * ntasks = 16 (MEDIUM) or 32 (LARGE, HUGE) for AlphaFold_v2.1.1
    * ntasks = 8 for AlphaFold_v2.1.2 and the following

Create SLURM script file

touch alphafold_monomer.sh

Then, copy-paste the example content (settings for SMALL monomeric protein) making sure to update the working directory path m_work and input file name input.fasta .

#!/bin/bash
#SBATCH --nodes=1                         # (required) number of requested nodes
#SBATCH --ntasks=8                        # (required) number of requested CPUs
#SBATCH --mem=180G                        # (required) memory allocation
#SBATCH --time=1:0:0                      # (required) runtime allocation
#SBATCH -p gpu                            # (required) type of GPU nodes
#SBATCH --gres=gpu:1                      # (required) number of requested GPUs (ONE only!)

#SBATCH --job-name="AF-TAG_pdb"           # (optional) job name visible in the queue
#SBATCH --output="out-TAG_pdb-%j"         # (optional) filename for standard output & errors

module load alphafold

### TIPS: use the full path for input and output, run sequences one by one.
m_work=/work/path_to/your/workdir
cd ${m_work}
time alphafold --fasta_paths=${m_work}/INPUTS/monomer.fasta --output_dir=${m_work}/OUTPUTS

Sumbit job to queue
sbatch alphafold_monomer.sh

Expected OUTPUTS

[]

Table of contents AlphaFold index Next