BioHPC Cloud:
: User Guide

Alphafold is currently installed on the rental GPU machines.  To use this software, you need to reserve time on one of these machines. (v2.3.2 is available on cbsugpu03-07.  v2.2.2 is available on cbsugpu02)

* Run time of alphafold:  cbsugpu07 is the fastest, cbsugpu-5-6 in between cbsugpu04 is the slowest. 

* To run alphafold multimer,  cbsugpu07 has 80GB of GPU RAM, cbsugpu05-6 have 48GB of GPU RAM.  More RAM allows you to work with larger protein complex.

* The script hard coded to set the following variables. 

TF_FORCE_UNIFIED_MEMORY=1
XLA_PYTHON_CLIENT_MEM_FRACTION=4.0

If you need to increase XLA_PYTHON_CLIENT_MEM_FRACTION to increase memory access, you can copy the /local/local_data/alphafold-2.3.2/run_singularity.py to your workdir, and modify the script by setting XLA_PYTHON_CLIENT_MEM_FRACTION to a bigger number.

To run Alphafold:

1. Create an output directory;

mkdir -p /workdir/$USER/alphafold_out

2. Copy your protein sequence fasta file(s) under /workdir/$USER

3. Start the command in a "screen" persistent session, from the directory /workdir/$USER (except between 30 min to 5 hours for a single protein, "multimer" mode might take up 1 to 2 days. To reduce run time, read this section  about multimer setting --num_multimer_predictions_per_model=1)

• In the option "--fasta_paths=xxx", xxx must be the fasta file name only, without directories.
• To run multimer, put all protein sequences in a single fasta file, and run with option "--model_preset=multimer";
• Output will be written to subfolders of "/workdir/$USER/alphafold_out" which you created in step 1.
• The option "max_template_date" allows you to exclude latest protein models in PDB to be used for prediction.
• More options are available, as described in https://github.com/deepmind/alphafold section Running AlphaFold, starting at step 4 (ignore the previous steps as these pertain to program installtion). For fasta_paths, either use full path or launch from directory where the fasta file is located. For example,

cd /workdir/$USER

#run version v2.3.2 (on cbsugpu03, cbsugpu04, cbsugpu05, cbsugpu06)
/local/local_data/alphafold-2.3.2/run_singularity.py --fasta_paths=T1050.fasta --max_template_date=2023-07-07 >& run.log &

#run version v2.2.2.(on cbsugpu02 only)
/programs/alphafold-2.2.2/alphafold_biohpc.sh --fasta_paths=T1050.fasta --max_template_date=2022-08-03 >& run.log

#run previous version 2.1.0 (on cbsugpu02 only)
/programs/alphafold/alphafold_biohpc.sh --fasta_paths=T1050.fasta --max_template_date=2022-08-03 >& run.log

4. If for any reason you need to terminate a run before it completes, press "ctrl-c"

QC Plots, see alphapickle

To extract ptm, iptm scores from pickle file of multimer:

See this page to evaluate scores.

## If you have several sequences, split them into fasta files with one sequence per file. In the command line, join these fasta files by ",". .  For example:

/local/local_data/alphafold-2.3.2/run_singularity.py --fasta_paths=prot1.fas,prot2.fas,prot3.fas --max_template_date=2022-07-08 >& my.log &

## if you have a large batch of sequences to process, you might want to split the CPU and GPU parts, and run on different BioHPC machine. You can optimize the two steps separately, e.g. put the hhblits databases on NVME disks  (There are some methods for separate the two steps, e.g. https://pythonrepo.com/repo/Zuricho-ParallelFold )