AlphaFold3
1. Introduction
This package provides an implementation of the inference pipeline of AlphaFold 3. See below for how to access the model parameters. You may only use AlphaFold 3 model parameters if received directly from Google. Use is subject to these terms of use.
Any publication that discloses findings arising from using this source code, the model parameters or outputs produced by those should cite the Accurate structure prediction of biomolecular interactions with AlphaFold 3 paper.
Please also refer to the Supplementary Information for a detailed description of the method.
AlphaFold 3 is also available at alphafoldserver.com for non-commercial use, though with a more limited set of ligands and covalent modifications.
If you have any questions, please contact the AlphaFold team at alphafold@google.com.
1.1 Local Installation
The CNCA team prepared a local installation of AlphaPhold3 using a container based on singularity or (appatainer) including the Genetic Database.
The Model Parameters is not included, users must request access filling a form and comply with the Google DeepMind terms of use at all times.
The local installation provide the AlphaFold3 version 3.0.1 over a container based on Ubuntu 24.04 distribution with cuda-12.6. The container is already prepared and is used throughtout a wrapper that accept all run_alphafold.py options; the Genetic Database location is already configured, users should provide the json path, model and output directories location.
The main resource target of AlphaFold3 is the GPU but the application can execute the data stage only on the CPU although the performance is substantially worst. The model inference stage requires a GPU, if you submit a job to a partition without a GPU, such as the hpc or fct partitions, then the option --norun-inference is added automatically to the run_alphafold.py script.
The user is free to prepare is own container as explained on https://github.com/google-deepmind/alphafold3/tree/main.
1.2 Environment
The environment is activate with command
$ module load alphaphold3/3.0.1
this will activate automatically a virtual environment ready to start the AlphaFold3 container throught the python script run_alphafold.py.
The command run_alphafold.py is a wrapper to the real script in the container and accept the same options. For example, you can get the run_alphafold.py help with command
$ run_alphafold.py --helpshort
1.3 Data Base Location
The Genetic Database is installed only on local disk of workernodes bellow the directory
/local/alphafold3
The environment variable ALPHAFOLD3_DATABASE points to this location.
The Genectic Database is not available on the user interfaces.
2. How to Run
2.1 Choose a working directory
$ mkdir ~/alphafold3_test
$ cd ~/alphafold3_test
2.2 Obtain the Model Parameters
The model parameters access is subject to Google DeepMind terms of use, please fill the form https://forms.gle/svvpY4u2jsHEwWYS6 and follow the instructions. Once you have access to the model parameters you will be responsible to keep the data private.
2.3 Create a submission script
Choose your favority editor and create a file, for example alphafold3.sh, with the following content:
$ emacs alphafold3.sh
#!/bin/bash
#SBATCH --partition=gpu
#SBATCH --gres=gpu
#SBATCH --nodes=1
#SBATCH --ntasks=8
#SBATCH --mem=64G
module purge
module load alphafold3
run_alphafold.py \
--json_path=fold_input.json \
--model_dir=./models \
--output_dir=./af_output
Create the models local directory and copy the model parameters into it:
$ mkdir models
You can place the model parameters on any other directory you just have to insert the appropriate directory location on the --model_dir option.
It is recommended to request at least 64GB of RAM but you can increase if necessary.
The output directory is created automatically if it does not exist.
2.4 Create the input JSON
You can use the example of section "Installation and Running Your First Prediction" of page https://github.com/google-deepmind/alphafold3?tab=readme-ov-file.
$ emacs fold_input.json
{
"name": "2PV7",
"sequences": [
{
"protein": {
"id": ["A", "B"],
"sequence" "GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG"
}
}
],
"modelSeeds": [1],
"dialect": "alphafold3",
"version": 1
}
2.5 Submit the job and check status
$ sbatch alphafold3.sh
Submitted batch job 22719987
$ squeue
JOBID PARTITION NAME USER ST TIME NODES CPUS TRES_PER_NODE NODELIST
22719987 gpu alphafold3 martinsj PD 0:00 1 8 gres/gpu
The squeue command list shows that the job is pending, or waiting for resources.
2.6 Check the running job
Eventually the job will run and the squeue command will show something like:
$ squeue
JOBID PARTITION NAME USER ST TIME NODES CPUS TRES_PER_NODE NODELIST
22719987 gpu alphafold3 martinsj R 1:25 1 8 gres/gpu hpc060
You can get more details about the job with command pestat:
$ pestat -j 22719987
Select only nodes with jobs in joblist=22719987
Hostname Partition Node Num_CPU CPUload Memsize Freemem GRES/ Joblist
State Use/Tot (MB) (MB) node JobId User GRES/job ...
hpc060 gpu mix 8 96 2.05* 512000 28466* gpu:t4:3(S:0-1) 22719987 martinsj
2.7 Check the finished job
On completion the job vanish from squeue list and you have the output on subdirectory "af_output" and the job standard output will be on file slurm-2271987.out:
$ ls -l
drwxr-x---+ 2 martinsj csys 4096 out 4 14:08 af_input
drwxr-x---+ 2 martinsj csys 4096 out 4 14:08 models
drwxrwx---+ 4 martinsj csys 4096 out 4 16:10 af_output
-rw-rw----+ 1 martinsj csys 424 out 4 14:29 alphafold3.sh
-rw-rw----+ 1 martinsj csys 46591 out 4 16:11 slurm-22719987.out
$ ls -l af_output/2PV7/
-rw-rw----+ 1 martinsj csys 3118163 Oct 4 16:11 2PV7_confidences.json
-rw-rw----+ 1 martinsj csys 7301275 Oct 4 16:11 2PV7_data.json
-rw-rw----+ 1 martinsj csys 409731 Oct 4 16:11 2PV7_model.cif
-rw-rw----+ 1 martinsj csys 147 Oct 4 16:11 2PV7_ranking_scores.csv
-rw-rw----+ 1 martinsj csys 332 Oct 4 16:11 2PV7_summary_confidences.json
drwxrwx---+ 2 martinsj csys 4096 Oct 4 16:11 seed-1_sample-0
drwxrwx---+ 2 martinsj csys 4096 Oct 4 16:11 seed-1_sample-1
drwxrwx---+ 2 martinsj csys 4096 Oct 4 16:11 seed-1_sample-2
drwxrwx---+ 2 martinsj csys 4096 Oct 4 16:11 seed-1_sample-3
drwxrwx---+ 2 martinsj csys 4096 Oct 4 16:11 seed-1_sample-4
-rw-rw----+ 1 martinsj csys 13036 Oct 4 16:11 TERMS_OF_USE.md
2.8 Get some job statistics
You can get some job statistics with commands:
$ seff 22719987
Job ID: 22719987
Cluster: production
User/Group: martinsj/csys
State: COMPLETED (exit code 0)
Nodes: 1
Cores per node: 8
CPU Utilized: 01:03:58
CPU Efficiency: 44.01% of 02:25:20 core-walltime
Job Wall-clock time: 00:18:10
Memory Utilized: 6.20 GB
Memory Efficiency: 6.46% of 96.00 GB
and
$ sacct -j 22719987
JobID JobName Partition Account AllocCPUS State ExitCode
------------ ---------- ---------- ---------- ---------- ---------- --------
22719987 alphafold+ gpu csys 8 COMPLETED 0:0
22719987.ba+ batch csys 8 COMPLETED 0:0
22719987.ex+ extern csys 8 COMPLETED 0:0
You can obtain more details with command
$ macct -j 22719987
3. Permance
It is recommended to request a NVIDIA A100 80GB GPU or a GPU with Compute Capatibily of 8.0 at least. The container is built with the NVIDIA A100 80GB in mind.
The CIRRUS provide three GPU types:
- NVIDIA A100 80GB
- NVIDIA Tesla V100s 32GB
- NVIDIA Tesla T4 16GB
The best performance will be obtained with the NVIDIA A100 80GB controller but this is the most requested resource and users may have to wait days to obtain one. The other controllers are less performant but more available and it may compensate to use the Tesla GPU's. As a reference, the example above completation times for each GPU was the following:
| Controller | Completion Time (hh:mm:ss) |
|---|---|
| NVIDIA A100 | 18:10 |
| NVIDIA Tesla V100s | 2:19:31 |
| NVIDIA Tesla T4 | 3:02:37 |
Users may request a specific controller modifying the option SBATCH --gres=gpu on the script to:
- #SBATCH --gres=gpu:a100
- #SBATCH --gres=gpu:v100s
- #SBATCH --gres=gpu:t4
to request the A100, or the Tesla V100s, or the Tesla T4, respectively. The Tesla GPU's controllers Compute Capabilities are less than 8.0 but it is possible to run alphafold3. The run_alphafold.py wrapper insert the right options on the container run_alphafold.py script in order to execute, although with reduced performance, on the Tesla GPU's.