Skip to main content

AlphaFold

Introduction​

AlphaFold is a groundbreaking AI system that is accelerating research in the field of bioinformatics. To use AlphaFold, the system first takes in a sequence of amino acids and then predicts the three-dimensional structure of a protein with extreme efficiency.

Read more on the AlphaFold official website.

This section on AlphaFold will guide you through using AlphaFold on Elja.

Getting started​

note

Due to NVIDIA compatibility issues, Elja now requires you to run AlphaFold in a Conda environment.

Setting up the Conda environment​

We start by initializing the conda environment, following the same steps as described in Conda:

$ module use /hpcapps/lib-mimir/modules/all 
$ module load Anaconda3/2023.09-0
$ conda config --add channels defaults
$ conda config --add channels bioconda
$ conda config --add channels conda-forge
$ conda config --set auto_activate_base false
$ conda init
$ bash # You can also log out and log in again.

Load AlphaFold​

Once conda is initialized and ready to use, we can load the AlphaFold module.

$ ml use /hpcapps/libbio-gpu/modules/all
$ ml load AlphaFold/2.3.1

Run AlphaFold on Elja​

note

AlphaFold will only run efficiently on GPU compute nodes. Be sure to specify a GPU partition when running your jobs.

To run AlphaFold on Elja, you can either use an interactive session or submit a batch job.

Starting an interactive session​

You can start an interactive session with the srun command on a GPU node. You can use the screen command or tmux to create a secondary terminal where your interactive session runs in the background.

$ srun --job-name "AlphaFold" --partition gpu-1xA100 --time 01:00:00 --pty bash
$ conda activate $env_path
$ run_alphafold.sh -d /AlphaFoldData/AlphaFold/data -o /hpcapps/source/alphafold_non_docker/dummy_test/ -f /hpcapps/source/alphafold_non_docker/example/query.fasta -t 2020-05-14

Running AlphaFold with SBATCH​

cat submit.slurm
#!/bin/bash
#SBATCH --mail-type=ALL
#SBATCH --mail-user=<MAIL>         # for example uname@hi.is
#SBATCH --nodes=1                  # number of nodes
#SBATCH --partition=gpu-1xA100
#SBATCH --time=1-00:00:00          # run for 1 day maximum
#SBATCH --output=slurm_job_output.log
#SBATCH --error=slurm_job_errors.log   # Logs if job crashes

module use /hpcapps/libbio-gpu/modules/all
module load AlphaFold/2.3.1
conda activate $env_path

# Run the command
run_alphafold.sh -d /AlphaFoldData/AlphaFold/data -o /hpcapps/source/alphafold_non_docker/dummy_test/ -f /hpcapps/source/alphafold_non_docker/example/query.fasta -t 2020-05-14

Additional Information​

AlphaFold Parameters​

When running AlphaFold using the run_alphafold.sh script, several parameters are available:

  • -d /AlphaFoldData/AlphaFold/data: Specifies the location of the AlphaFold database (required)
  • -o <output_dir>: Directory where results will be saved
  • -f <fasta_file>: Path to the FASTA file containing the protein sequence
  • -t <max_template_date>: Maximum template release date (YYYY-MM-DD)

For a complete list of parameters, refer to the AlphaFold documentation.

Interpreting Results​

AlphaFold generates several files for each prediction:

  • PDB files containing the predicted structures
  • JSON files with confidence metrics
  • Visualization files for examining the quality of predictions

The primary metric for evaluating prediction quality is the pLDDT score (predicted Local Distance Difference Test), which ranges from 0 to 100, with higher values indicating higher confidence.

Troubleshooting​

Common issues when running AlphaFold on Elja:

  1. CUDA errors: Ensure you're using the correct GPU partition
  2. Memory limitations: Large proteins may require more GPU memory; adjust batch sizes if needed
  3. Environment errors: Verify that the Conda environment is properly activated

For additional help, contact the Elja support team.