Introduction to AlphaFold for 3D Protein

Structure Prediction on Grace

Texas A&M Research Computing Symposium
May 25, 2022

High Performance Research Computing | hprc.tamu.edu 1

AlphaFold for 3D Protein Structure
Prediction on Grace
● Resources and Limitations
● Database Files
● Running AlphaFold
○ Google Colab
○ ChimeraX + Google Colab
○ Grace GPU or non-GPU nodes
● Visualization of Results
○ job resource usage
○ view predictions in ChimeraX
○ plotting pLDDT values
● Alternative Workflows
https://hprc.tamu.edu/wiki/SW:AlphaFold

High Performance Research Computing | hprc.tamu.edu 2

 Resource Limitations
● AlphaFold
○ currently AlphaFold can only utilize one GPU
○ about 90% of processing is done on CPU when using
DeepMind's workflow
● AlphaFold on HPRC Grace
○ sometimes GPU not detected on certain nodes
● AlphaFold in Google Colab (web browser or ChimeraX app)
○ no guarantee of available resources in Colab
○ runs as a Jupyter notebook on Google Colab cloud servers
■ 12GB RAM max
■ a notebook can run for up to 12 hours per day
● 24 hours per day with Colab Pro ($9.99/month)
■ not suitable for large predictions

High Performance Research Computing | hprc.tamu.edu 3

 AlphaFold Databases on Grace
/scratch/data/bio/alphafold/2.2.0
├── bfd ├── pdb70
│ ├── [1.4T] bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_a3m.ffdata │ ├── [ 410] md5sum
│ ├── [1.7G] bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_a3m.ffindex │ ├── [ 53G] pdb70_a3m.ffdata
│ ├── [ 16G] bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_cs219.ffdata │ ├── [2.0M] pdb70_a3m.ffindex
│ ├── [1.6G] bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_cs219.ffindex │ ├── [6.6M] pdb70_clu.tsv
│ ├── [304G] bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_hhm.ffdata │ ├── [ 21M] pdb70_cs219.ffdata
│ └── [124M] bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt_hhm.ffindex │ ├── [1.5M] pdb70_cs219.ffindex
├── mgnify │ ├── [3.2G] pdb70_hhm.ffdata
│ └── [ 64G] mgy_clusters_2018_12.fa │ ├── [1.8M] pdb70_hhm.ffindex
├── mgnify_2019_05 │ └── [ 19M] pdb_filter.dat
│ └── [271G] mgy_proteins.fa ├── pdb_mmcif
├── params │ ├── [9.4M] mmcif_files
│ ├── [ 18K] LICENSE │ └── [141K] obsolete.dat
│ ├── [356M] params_model_1_multimer_v2.npz ├── pdb_seqres
│ ├── [356M] params_model_1.npz │ └── [218M] pdb_seqres.txt
│ ├── [356M] params_model_1_ptm.npz ├── small_bfd
│ ├── [356M] params_model_2_multimer_v2.npz │ └── [ 17G] bfd-first_non_consensus_sequences.
│ ├── [356M] params_model_2.npz ├── uniclust30
│ ├── [356M] params_model_2_ptm.npz │ ├── [ 87G] uniclust30_2018_08
│ ├── [356M] params_model_3_multimer_v2.npz │ └── [ 206G] uniclust30_2021_03
│ ├── [354M] params_model_3.npz ├── uniprot
│ ├── [355M] params_model_3_ptm.npz │ └── [104G] uniprot.fasta
│ ├── [356M] params_model_4_multimer_v2.npz └── uniref90
│ ├── [354M] params_model_4.npz └── [ 62G] uniref90.fasta
│ ├── [355M] params_model_4_ptm.npz
│ ├── [356M] params_model_5_multimer_v2.npz
│ ├── [354M] params_model_5.npz total size: 2.6TB
│ └── [355M] params_model_5_ptm.npz number of files: 183,000+

High Performance Research Computing | hprc.tamu.edu 4

 Resources for Running AlphaFold

● Run as a Jupyter Notebook on Google Colab in web browser

● Run as a Jupyter Notebook on Google Colab in ChimeraX
○ ChimeraX Interactive App on HPRC Grace Portal
■ https://portal-grace.hprc.tamu.edu
● Run as a Slurm job script on Grace

https://hprc.tamu.edu/wiki/SW:AlphaFold

High Performance Research Computing | hprc.tamu.edu 5

ColabFold AlphaFold2 Jupyter Notebook
Enter an amino acid sequence
NLYIQWLKDGGPSSGRPPPS

High Performance Research Computing | hprc.tamu.edu 6

 ChimeraX
● Can be used to visualize protein structures
● Can be launched using the Grace portal
○ portal-grace.hprc.tamu.edu
● Can be used to run AlphaFold using the daily build version (2022.02.22+)
○ uses Google Colab with limited resources

High Performance Research Computing | hprc.tamu.edu 7

● Launch a ChimeraX job on the Grace
portal portal-grace.hprc.tamu.edu
○ select Node Type: ANY
○ AlphaFold runs on Google Colab
GPUs so we can use a non-GPU
for running ChimeraX
● ChimeraX will be used later for the
following
○ run AlphaFold in Google Colab
○ visualize results from an
AlphaFold job on Grace

High Performance Research Computing | hprc.tamu.edu 8

 Running AlphaFold on Grace
● Can be run as a job script requesting one GPU
● shared databases are available: 2.6TB total size
benchmark monomer_ptm 98 aa multimer 98 & 73 aa

A100 1 hour 26 minutes 4 hours 49 minutes

RTX 6000 2 hours 39 minutes 4 hours 44 minutes

T4 2 hours 35 minutes 4 hours 45 minutes

CPU only 2 hours 50 minutes 6 hours 11 minutes

● Can be run in ChimeraX from the Grace portal

○ using the ChimeraX AlphaFold option
○ all processing done on Google cloud servers
■ monomer_ptm (98 aa) on GPU = 1 hour 14 minutes
High Performance Research Computing | hprc.tamu.edu 9
 Finding AlphaFold template job scripts
using GCATemplates on Grace
● Genomic Computational Analysis
Templates have example input data so
you can run the script for demo purposes
mkdir $SCRATCH/af2demo
cd $SCRATCH/af2demo
gcatemplates
● Type s for search then enter alphafold to
search for the alphafold 2.2.0 template
script and select the reduced_dbs script
● Review the script and submit the job
script which takes about 30 minutes to
complete
sbatch run_alphafold_2.2.0_reduced_dbs_monomer_ptm_grace.sh

High Performance Research Computing | hprc.tamu.edu 10

 Monitoring AlphaFold GPUs Found
● Check to make sure that AlphaFold can detect GPUs
○ wait a few minutes after the job starts
○ search for the text "No GPU/TPU found, falling back to CPU."
■ grep CPU stderr*
● If the job did not detect GPUs
○ find the compute node name in the NodeList column
■ sacct
sacct-j-j jobID
jobID
○ cancel the job
■ scancel jobID
scancel jobID
○ add a line in your job script to ignore the compute node
■ #SBATCH --exclude=g016
○ submit your updated job script
○ send an email to the HPRC helpdesk with the node name
■ [email protected]

High Performance Research Computing | hprc.tamu.edu 11

 AlphaFold with
ChimeraX + Google Colab

High Performance Research Computing | hprc.tamu.edu 12

Maximize ChimeraX Window

Right click and select

Maximize if the
ChimeraX window is off
the screen

High Performance Research Computing | hprc.tamu.edu 13

 AlphaFold with ChimeraX
● Launch ChimeraX using the HPRC Grace portal
● Select the AlphaFold Structure Prediction option

High Performance Research Computing | hprc.tamu.edu 14

 AlphaFold with ChimeraX
● Enter an amino acid sequence
NLYIQWLKDGGPSSGRPPPS
○ or paste in Clipboard first
then paste in Sequence
field
● Click Predict
● A Google Colab page will
start and prompt you for
your Google login
● Login to your Google
account to begin processing
● Prediction completes in
about 1 hour
● Not ideal for large prediction
jobs
High Performance Research Computing | hprc.tamu.edu 15
AlphaFold Grace Job Scripts

High Performance Research Computing | hprc.tamu.edu 16

 Example AlphaFold Job Script
#!/bin/bash
#SBATCH --job-name=alphafold # job name
#SBATCH --time=2-00:00:00 # max job run time dd-hh:mm:ss

● multimer #SBATCH --ntasks-per-node=1

#SBATCH --cpus-per-task=24
#
#
tasks (commands) per compute node
CPUs (threads) per command

○ dbs in red required for

#SBATCH --mem=180G # total memory per node
#SBATCH --gres=gpu:a100:1 # request 1 A100 GPU
#SBATCH --output=stdout.%x.%j # save stdout to file

multimer #SBATCH --error=stderr.%x.%j # save stderr to file

AlphaFold can only use one

module purge
● export SINGULARITYENV_TF_FORCE_UNIFIED_MEMORY=1

GPU so reserve half the CPU

export SINGULARITYENV_XLA_PYTHON_CLIENT_MEM_FRACTION=4.0

DOWNLOAD_DIR=/scratch/data/bio/alphafold/2.2.0

and memory resources so # run gpustats in the background (&) to monitor gpu usage in order to create a graph later

another job can use the other

gpustats &

singularity exec --nv /sw/hprc/sw/bio/containers/alphafold/alphafold_2.2.0.sif \

GPU python /app/alphafold/run_alphafold.py \

--use_gpu_relax \

○ Grace compute nodes

--data_dir=$DOWNLOAD_DIR \
--uniref90_database_path=$DOWNLOAD_DIR/uniref90/uniref90.fasta \
--mgnify_database_path=$DOWNLOAD_DIR/mgnify/mgy_clusters_2018_12.fa \

have 360GB of available --uniclust30_database_path=$DOWNLOAD_DIR/uniclust30/uniclust30_2021_03/UniRef30_2021_03 \

--bfd_database_path=$DOWNLOAD_DIR/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \

memory and 48 cores

--model_preset=multimer \
--pdb_seqres_database_path=$DOWNLOAD_DIR/pdb_seqres/pdb_seqres.txt \
--uniprot_database_path=$DOWNLOAD_DIR/uniprot/uniprot.fasta \
--template_mmcif_dir=$DOWNLOAD_DIR/pdb_mmcif/mmcif_files \
--obsolete_pdbs_path=$DOWNLOAD_DIR/pdb_mmcif/obsolete.dat \
--max_template_date=2022-1-1 \
--db_preset=full_dbs \
--output_dir=out_alphafold \
--fasta_paths=$DOWNLOAD_DIR/example_data/T1083_T1084_multimer.fasta

# run gpustats to create a graph of gpu usage for this job

gpustats

High Performance Research Computing | hprc.tamu.edu 17

 Example AlphaFold Job Script
#!/bin/bash
#SBATCH --job-name=alphafold # job name
#SBATCH --time=2-00:00:00 # max job run time dd-hh:mm:ss

● monomer #SBATCH --ntasks-per-node=1

#SBATCH --cpus-per-task=24
#
#
tasks (commands) per compute node
CPUs (threads) per command

○ dbs in red required for

#SBATCH --mem=180G # total memory per node
#SBATCH --gres=gpu:a100:1 # request 1 A100 GPU
#SBATCH --output=stdout.%x.%j # save stdout to file

monomer #SBATCH --error=stderr.%x.%j # save stderr to file

monomer_ptm
module purge
● export SINGULARITYENV_TF_FORCE_UNIFIED_MEMORY=1

○ will produce pTM scores

export SINGULARITYENV_XLA_PYTHON_CLIENT_MEM_FRACTION=4.0

DOWNLOAD_DIR=/scratch/data/bio/alphafold/2.2.0

that can be graphed # run gpustats in the background (&) to monitor gpu usage in order to create a graph later

using AlphaPickle
gpustats &

singularity exec --nv /sw/hprc/sw/bio/containers/alphafold/alphafold_2.2.0.sif \

● AlphaFold can only use one python /app/alphafold/run_alphafold.py \

--use_gpu_relax \

GPU so reserve half the CPU

--data_dir=$DOWNLOAD_DIR \
--uniref90_database_path=$DOWNLOAD_DIR/uniref90/uniref90.fasta \
--mgnify_database_path=$DOWNLOAD_DIR/mgnify/mgy_clusters_2018_12.fa \

and memory resources so --uniclust30_database_path=$DOWNLOAD_DIR/uniclust30/uniclust30_2021_03/UniRef30_2021_03 \

--bfd_database_path=$DOWNLOAD_DIR/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \

another job can use the

--model_preset=monomer \
--pdb70_database_path=$DOWNLOAD_DIR/pdb70/pdb70 \
--template_mmcif_dir=$DOWNLOAD_DIR/pdb_mmcif/mmcif_files \

other GPU --obsolete_pdbs_path=$DOWNLOAD_DIR/pdb_mmcif/obsolete.dat \

--max_template_date=2022-1-1 \
--db_preset=full_dbs \
--output_dir=out_alphafold \
--fasta_paths=$DOWNLOAD_DIR/example_data/T1083.fasta

# run gpustats to create a graph of gpu usage for this job

gpustats

High Performance Research Computing | hprc.tamu.edu 18

 Example AlphaFold Job Script
#!/bin/bash
#SBATCH --job-name=alphafold # job name
#SBATCH --time=2-00:00:00 # max job run time dd-hh:mm:ss

● monomer + reduced_dbs #SBATCH --ntasks-per-node=1

#SBATCH --cpus-per-task=24
#
#
tasks (commands) per compute node
CPUs (threads) per command

dbs in red required for

#SBATCH --mem=180G # total memory per node
● #SBATCH --gres=gpu:a100:1
#SBATCH --output=stdout.%x.%j
#
#
request 1 A100 GPU
save stdout to file

monomer + reduced_dbs #SBATCH --error=stderr.%x.%j # save stderr to file

small_bfd_database is a
module purge
● export SINGULARITYENV_TF_FORCE_UNIFIED_MEMORY=1

subset of BFD and is

export SINGULARITYENV_XLA_PYTHON_CLIENT_MEM_FRACTION=4.0

DOWNLOAD_DIR=/scratch/data/bio/alphafold/2.2.0

generated by taking the first # run gpustats in the background (&) to monitor gpu usage in order to create a graph later

non-consensus sequence
gpustats &

singularity exec --nv /sw/hprc/sw/bio/containers/alphafold/alphafold_2.2.0.sif \

from every cluster in BFD python /app/alphafold/run_alphafold.py \

--use_gpu_relax \

AlphaFold can only use one

--data_dir=$DOWNLOAD_DIR \
● --uniref90_database_path=$DOWNLOAD_DIR/uniref90/uniref90.fasta \
--mgnify_database_path=$DOWNLOAD_DIR/mgnify/mgy_clusters_2018_12.fa \

GPU so reserve half the CPU --small_bfd_database_path=$DOWNLOAD_DIR/small_bfd/bfd-first_non_consensus_sequences.fasta \

--model_preset=monomer \

and memory resources so

--pdb70_database_path=$DOWNLOAD_DIR/pdb70/pdb70 \
--template_mmcif_dir=$DOWNLOAD_DIR/pdb_mmcif/mmcif_files \
--obsolete_pdbs_path=$DOWNLOAD_DIR/pdb_mmcif/obsolete.dat \

another job can use the other --max_template_date=2022-1-1 \

--db_preset=reduced_dbs \
--output_dir=out_alphafold \
GPU --fasta_paths=$DOWNLOAD_DIR/example_data/T1083.fasta

# run gpustats to create a graph of gpu usage for this job

gpustats

High Performance Research Computing | hprc.tamu.edu 19

 Unified Memory
#!/bin/bash
#SBATCH --job-name=alphafold # job name
#SBATCH --time=2-00:00:00 # max job run time dd-hh:mm:ss
#SBATCH --ntasks-per-node=1 # tasks (commands) per compute node
#SBATCH --cpus-per-task=24 # CPUs (threads) per command
#SBATCH --mem=180G # total memory per node
#SBATCH --gres=gpu:a100:1 # request 1 A100 GPU
#SBATCH --output=stdout.%x.%j # save stdout to file
#SBATCH --error=stderr.%x.%j # save stderr to file

export SINGULARITYENV_TF_FORCE_UNIFIED_MEMORY=1
export SINGULARITYENV_XLA_PYTHON_CLIENT_MEM_FRACTION=4.0

● unified memory can be used to request more than just the total GPU
memory for the JAX step in AlphaFold
○ A100 GPU has 40GB memory
○ GPU total memory (40) * XLA_PYTHON_CLIENT_MEM_FRACTION (4.0)
○ XLA_PYTHON_CLIENT_MEM_FRACTION default = 0.9
● this example script has 160 GB of unified memory
○ 40 GB from A100 GPU + 120 GB DDR from motherboard

High Performance Research Computing | hprc.tamu.edu 20

 AlphaFold Results
Visualization

High Performance Research Computing | hprc.tamu.edu 21

 Review GPU usage for a Job
The gpustats command monitors GPU resource usage and create a graph

#!/bin/bash
#SBATCH --job-name=my_gpu_job
#SBATCH --time=1-00:00:00
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=24
#SBATCH --mem=180G
#SBATCH --gres=gpu:a100:1
#SBATCH --output=stdout.%x.%j
#SBATCH --error=stderr.%x.%j

module purge

# run gpustats in the background (&) to monitor gpu usage

gpustats &

my_alphafold_command

# run gpustats to create a graph of gpu usage for this job

gpustats

When the job is complete, login with ssh -X option and eog stats_gpu.3411850.png
view graph of GPU usage stats using the eog command

High Performance Research Computing | hprc.tamu.edu 22

Review CPU usage for a Job
The seff command displays CPU resource usage

seff 3411850

Job ID: 3411850

Cluster: grace
User/Group: netid/netid
State: COMPLETED (exit code 0)
Nodes: 1
Cores per node: 24
CPU Utilized: 06:24:58
CPU Efficiency: 2.70% of 9-21:48:24 core-walltime
Job Wall-clock time: 09:54:31
Memory Utilized: 54.59 GB
Memory Efficiency: 30.33% of 180.00 GB

usage stats are not accurate until the job is complete

High Performance Research Computing | hprc.tamu.edu 23

Visualize Results with ChimeraX

Click the minimize

button to return to
ChimeraX or click
Reconnect then
minimize

High Performance Research Computing | hprc.tamu.edu 24

Visualize AlphaFold Google Colab Results
with ChimeraX

High Performance Research Computing | hprc.tamu.edu 25

 View PDB Structure if Available

type open and the

protein name 1L2Y

High Performance Research Computing | hprc.tamu.edu 26

Visualize AlphaFold Grace Results
with ChimeraX

High Performance Research Computing | hprc.tamu.edu 27

AlphaFold Confidence Metrics

High Performance Research Computing | hprc.tamu.edu 28

 Visualize AlphaFold .pkl Results

> 90 = Very high

70 - 90 = Confident
50 - 70 = Low
< 50 = Very low

eog output_dir/protein_dir/ranked_0_pLDDT.png

you may get different results compared to the image above when using reduced_dbs

High Performance Research Computing | hprc.tamu.edu 29

Visualize AlphaFold PAE Results (monomer_ptm)
● Low Predicted Aligned Error
(PAE) value has higher
confidence of accuracy
● Must use monomer_ptm or
Aligned residue

multimer as model_preset to
create PAE image
● The colour at position (x, y)
indicates AlphaFold’s
expected position error at
residue x, when the predicted
and true structures are
aligned on residue y.
Scored residue

eog output_dir/protein_dir/ranked_0_PAE.png https://alphafold.ebi.ac.uk

High Performance Research Computing | hprc.tamu.edu 30

 Evaluating Models

ranked_0 ranked_1 ranked_2 ranked_3 ranked_4

"plddts": {
"model_1_ptm_pred_0": 94.16585478746399,
"model_2_ptm_pred_0": 94.64120852328334,
"model_3_ptm_pred_0": 89.94980057627299,
"model_4_ptm_pred_0": 77.53515668415058,
see which model has the top rank based on pLDDT score },
"model_5_ptm_pred_0": 88.40610380463586

"order": [
"model_2_ptm_pred_0", ranked_0
"model_1_ptm_pred_0",
"model_3_ptm_pred_0",
cat out_IL2Y_reduced_dbs_monomer_ptm/IL2Y/ranking_debug.json "model_5_ptm_pred_0",
"model_4_ptm_pred_0"
]

High Performance Research Computing | hprc.tamu.edu 31

AlphaFold Workflow
Alternatives

High Performance Research Computing | hprc.tamu.edu 32

 ParallelFold
● ParallelFold (ParaFold) breaks the AlphaFold workflow into two steps
○ processing of the three CPU steps in parallel
○ processing of the GPU step
● The parallel portion for CPU steps is not implemented yet resulting in
similar or longer runtimes than the DeepMind approach
○ the first three CPU steps, jackhammer, jackhammer and HHblits
are supposed to run as three separate processes in parallel but
currently this parallelization step is not implemented yet
○ when the CPU processing is parallelized, it will require 21 cores
■ 8 cores for each of the two jackhammer steps
■ 5 cores for the HHblits step.
● These steps may be implemented in parallel in AlphaFold soon
https://github.com/Zuricho/ParallelFold

High Performance Research Computing | hprc.tamu.edu 33

Databases and References

High Performance Research Computing | hprc.tamu.edu 34

DeepMind and EMBL’s
European Bioinformatics
Institute (EMBL-EBI) have
partnered to create
AlphaFold DB to make
these predictions freely
available to the scientific
community.

High Performance Research Computing | hprc.tamu.edu 35

 References

Arnold, M. J. (2021) AlphaPickle doi.org/10.5281/zenodo.5708709

High Performance Research Computing | hprc.tamu.edu 36