AlphaFold#
AlphaFold can accurately predict 3D models of protein structures and is accelerating research in nearly every field of biology.
- Deepmind
AlphaFold is installed on the HPC Cluster using Singularity images. The code version we use is an alternative to the native docker container they provide.
Reference Files#
AlphaFold downloaded reference files are located at /hpc/refdata/alphafold/2.3. This path is set for you in the alphafold software module by default when you use the -d $DOWNLOAD_DIR flag.
Example AlphaFold interactive job#
# start interactive job
srun --job-name=alphafold_test --ntasks=1 --cpus-per-task=12 --mem=30gb --time=8:00:00 --gres=gpu:1 --gpu_cmode=shared --pty bash
# load required modules on the gpu node
module load alphafold/2.3.2
module load python/3.9.1
# tell alphafold about the GPU
export NVIDIA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES
# start alphafold software, replace netid with your username
run_alphafold.py -d $DOWNLOAD_DIR --cpus 12 -f /scratch/g/PI_NetID/alphatest/some.fasta -t 2020-05-14 -o /scratch/g/PI_NetID/alphatest/output
The example above is run as a normal interactive SLURM job. SLURM will assign a GPU node for interactive use. The alphafold software environment is loaded using module load alphafold/2.3.2 command. Please make sure to insert your username for PI_NetID.
Example AlphaFold batch job#
Create a file /scratch/g/PI_NetID/alphatest/alphafold.slurm containing the following. Replace PI_NetID with your PI's Username.
#!/bin/bash
#SBATCH --job-name=alphafold_test
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=12
#SBATCH --mem=30gb
#SBATCH --time=08:00:00
#SBATCH --output=%x-%j.out
#SBATCH --gres=gpu:1
#SBATCH --gpu_cmode=shared
#SBATCH --partition=gpu
module load alphafold/2.3.2
module load python/3.9.1
export NVIDIA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES
run_alphafold.py -d $DOWNLOAD_DIR --cpus 12 -f /scratch/g/PI_NetID/alphatest/some.fasta -t 2020-05-14 -o /scratch/g/PI_NetID/alphatest/output
Submit the job:
cd /scratch/g/PI_NetID/alphatest && sbatch alphafold.slurm
The example above is a job request for 1 node, 12 processors, and 1 gpu.
GPU Compute Mode
The #SBATCH --gpu_cmode=shared slurm header is required or your job will fail halfway through. Alphafold 2.3.2 doesn't parallelize on GPUs so you should only run on 1 GPU.
Help#
Specific help information is available on GitHub. Additional command-line help can be printed with module help alphafold/2.3.2.