AlphaFold 3 is an advanced artificial intelligence system developed by Google DeepMind and Isomorphic Labs that predicts the 3D structures and interactions of a wide range of biological molecules, including proteins, DNA, RNA, and ligands (small drug-like molecules). It represents a significant advancement over previous versions by accurately modeling complex molecular interactions, offering a comprehensive view of cellular systems that accelerates research in areas like drug discovery and the understanding of biological functions.

Model Parameters Terms of Use

Google DeepMind has revised the licensing terms for AlphaFold software with the release of version 3. The source code continues to be open source, but the model parameters—derived from training the AlphaFold model and necessary for performing molecular structure predictions—are now distributed separately under their own usage terms. HKU users interested in running AlphaFold 3 should review and understand the specific terms of use applicable to the model parameters::

User may request a copy of the model parameters by filling out and submitting this form. Approval of the request is entirely at the discretion of Google DeepMind. 

Note: To comply with the terms of use, each AlphaFold 3 user in HKU must request, download, and use their own personal copy of the model parameters.

Follow these instructions to complete the form:

  1. Review and understand the section titled Key things to know when using the AlphaFold 3 model parameters and output.
  2. In the first email field labeled “Email”, enter your HKU email address.
  3. In the second email field labeled “Google account email address”, provide a Gmail address rather than a HKU email address. If you do not have a Gmail account, go to gmail.com and create one.
  4. In the field labeled “URL of public-facing website for non-commercial organization”, enter https://www.hku.hk.
  5. Once all fields on the first page are completed, click the Next button.
  6. On the second page, answer “No” to the question: “Do you intend to provide access to the AlphaFold 3 model parameters to other researchers within your non-commercial organization (e.g., as part of a centrally managed computing cluster)?”
  7. Click Next, review the terms on the final page, and if you agree, complete and submit the form.

Using AlphaFold 3 with GPU nodes on HPC2021

Preparation

# Create the folder for AlphaFold 3 Model
mkdir -p ~/af3/models

# Upload the model file downloaded from Google to ~/af3/models and decompress it
unzstd af3.bin.zst 

# Create the folder for input to AlphaFold
mkdir ~/af3/input
cp /share1/alphafold/3.0.1/sample/input/fold_input.json ~/af3/input

# Create the output folder
mkdir ~/af3/output

Put the following in a SLURM script and submit it as a batch job to run with a single GPU

(AlphaFold 3 does not support multi GPU)

#!/bin/bash
#SBATCH --ntasks=16
#SBATCH --nodes=1
#SBATCH --partition=l40s
#SBATCH --qos=gpu
#SBATCH --gres=gpu:1

ml alphafold/3.0.1

singularity exec \
--nv \
--bind $HOME/af3/input:/root/af_input \
--bind $HOME/af3/output:/root/af_output \
--bind $HOME/af3/models:/root/models \
--bind /software/alphafold/db.v3:/root/public_databases \
/share1/alphafold/3.0.1/af3.0.1.sif \
run_alphafold.py \
--flash_attention_implementation=triton \
--json_path=/root/af_input/fold_input.json \
--model_dir=/root/models \
--db_dir=/root/public_databases \
--output_dir=/root/af_output \
--jackhmmer_n_cpu=16 \
--nhmmer_n_cpu=16